PyPI · #2987 most downloaded on PyPI

openmed

2.5.01.8M downloads/momaziyarpanahi/openmed ↗New releases in your daily digest, changelog included.

Local-first SDK for clinical extraction and de-identification workflows on hardware you control.

Last release 19 days ago
15 Sep 2026
Ships on a steady schedule
a new release about every 2 weeks
Nearly every release is documented
notes for 41 of 44 stable releases
Nothing withdrawn
no release was ever pulled
1 years old
50 releases · first in 2025

Release timeline

50 releases since 2025

20252026

One column per month.

Releases

1.7.01 Jul 2026Security2 sources agree
root SECURITY.md and private vulnerability reporting guidance
OpenMed 1.7.0 is the multimodal de-identification and evaluation-depth release.

This release turns the v1.6 privacy assurance foundation into a broader local-first clinical data platform: multimodal document intake, OCR adapters, structured clinical extraction, FHIR/HL7/CDA/CSV de-identification, streaming and batch service controls, richer policy profiles, and typed clients for Python, Swift, and TypeScript.

It also deepens the evidence layer around releases: benchmark scorecards, threshold sweeps, leakage heatmaps, membership-inference probes, k-anonymity metrics, utility-loss reports, release-gate previews, evidence bundles, SBOMs, reproducible locks, and stronger supply-chain checks.

Release date: 2026-07-01.

Highlights

Added multimodal document primitives, source spans, OCR adapters, image/PDF redaction, Markdown/AsciiDoc offset extraction, metadata scrubbing, JSONL chat-log de-identification, and language-aware OCR configuration.

Added de-identification adapters for CDA/C-CDA XML, HL7 v2, CSV/TSV, FHIR $de-identify, FHIR Bulk NDJSON, deterministic FHIR bundles, OperationOutcome, Provenance, AuditEvent, and CodeableConcept validation/provenance helpers.

Added clinical extraction and normalization for lab values, vital signs, medication sigs, problem lists, summary cards, microbiology concepts, dermatology and ophthalmology domains, clinical concept labels, assertion context, status vocabulary, and clinical-term protection.

Expanded language and locale coverage with Indonesian, Thai, Hebrew RTL, Polish PESEL, Korean RRN, Unicode script segmentation, locale national-ID providers, and deterministic locale PHI generation.

Added runtime de-identification features including typed AnalyzeResult, DeidentificationResult.to_dataframe(), redaction preview diffs, cross-document surrogate vaults, patient-keyed date shifting, format-preserving redaction, minimum-necessary strength selection, custom recognizers, streaming incremental de-identification, explain traces, section stamping, and per-document risk budgets.

Added new CLI and SDK surfaces: openmed deid, openmed fhir bundle, openmed models recommend, openmed models diff, openmed policy diff, openmed doctor, openmed gates preview, openmed gates bundle, openmed audit, openmed risk, model-card previews, a typed Python REST client, a TypeScript REST client, and Swift policy-profile support.

Added REST service operations for model warm pools, dynamic batching, request coalescing, rate and concurrency limits, readiness/liveness split, graceful shutdown, trusted hosts, CORS configuration, Prometheus-style metrics, Docker Compose, and Kafka streaming de-identification.

Added evaluation and release evidence tooling for DrugProt, public biomedical NER, i2b2, clinical PHI manifests, dataset cards, fixture coverage, per-section recall, result caching, fairness, robustness, error analysis, leakage heatmaps, membership inference, linkage attacks, k-anonymity/l-diversity/t-closeness, audit diffs, model scorecards, threshold sweeps, flaky-run detection, paired significance, calibration reliability, utility loss, policy compliance, benchmark history diffs, nano-tier certification, and risk dashboards.

Added model and backend work for Laneformer MLX-LM, MLX INT4 certification, Core ML INT8 export, AWQ/GPTQ recipes, bitsandbytes 4-bit loading, FlashAttention/SDPA/eager selection, PyTorch MPS tuning, ONNX/WebGPU and Transformers.js exports, tokenizer caching, Mode-A distillation, DAPT corpus assembly, DirectID, hard-negative sampling, and span-relation graph decoding.

Added security and release engineering coverage: root responsible disclosure policy, breach runbook and report template, CycloneDX SBOM generation, reproducible-lock workflow, lockfile drift gate, GitHub Actions ref validation, manifest schema validation, dependency updates, doctest-backed examples, and PHI-safe diagnostics.

The Big Picture

OpenMed v1.7.0 is about broadening what can be safely de-identified and proving more of the surrounding workflow.

The v1.6 release established policy-aware spans, audit reports, risk scoring, and release gates. v1.7 applies that foundation to more input types and operational paths: scanned documents, PDFs, Markdown and AsciiDoc, structured tables, HL7 v2, CDA/C-CDA XML, FHIR operations, bulk NDJSON, chat logs, Kafka streams, and REST clients.

The evaluation story also moves from single aggregate scores toward release-grade evidence: per-language leakage heatmaps, per-section recall, dataset cards, fixture coverage, model scorecards, threshold sweeps, fairness and robustness reports, utility-loss metrics, and evidence bundles that make gate decisions easier to inspect.

Multimodal And Document Redaction

The new multimodal layer gives OpenMed a shared contract for document ingestion and source-preserving redaction.

This release adds:

openmed.multimodal primitives for document content, source spans, handlers, and lazy registration

image redaction and OCR result contracts

Tesseract, PaddleOCR, EasyOCR, and docTR engine adapters

OCR language configuration for non-English scans

PDF coordinate projection for detected spans

Markdown and AsciiDoc text extraction with character-offset maps

metadata scrubbing and verification for images, PDFs, and DOCX files

JSONL chat-log de-identification with role and turn structure preservation

The result is a cleaner path from scanned or mixed-format clinical material to audited text, coordinates, redactions, and PHI-safe summaries.

Interoperability And Structured Health Data

v1.7.0 adds a large set of healthcare interchange helpers.

New or expanded interop paths include:

CDA/C-CDA XML de-identification

HL7 v2 message parsing and field-level redaction

CSV/TSV PHI column classification and redaction manifests

FHIR $de-identify operation wrappers

FHIR Bulk NDJSON de-identification

deterministic FHIR Bundle assembly and stable urn:uuid references

FHIR OperationOutcome generation

FHIR Provenance and AuditEvent emission from signed audit reports

CodeableConcept builders, text fallback checks, and code-system version pinning

PHILTER, pyDeid, GLiNER-BioMed, LangChain, and spaCy adapter surfaces

These additions make OpenMed more useful in pipelines that already speak health-data formats instead of plain text alone.

Clinical Extraction And Context

Clinical extraction is now broader and more context-aware.

This release adds or expands:

dermatology and ophthalmology zero-shot domains

microbiology labels and routing metadata

lab reference range parsing and flag handling

vital signs extraction

medication sig normalization

problem-list deduplication and status reconciliation

clinical summary cards without PHI

clinical concept canonical labels

clinical assertion context records

negation, section-aware ConText assertions, and sentence-bounded cue scope

clinical term protection to reduce false PII hits on protected clinical vocabulary

status vocabulary normalization for substance, employment, and living-status language

Together these changes reduce over-redaction risk and improve downstream structured clinical outputs.

Language, Locale, And Identifier Coverage

v1.7.0 expands multilingual PII coverage and locale correctness.

New coverage includes:

Indonesian PII with NIK validation

Thai PII patterns and fixtures

Hebrew RTL PII handling

Polish PESEL validation

South Korean RRN validation

Unicode script detection and mixed-script segmentation

locale national-ID provider registry

per-language surrogate coherence regression coverage

locale-aware date and number normalization

deterministic locale PHI generation for training and evaluation

Runtime, Service, And Clients

The runtime and service layers are more production-oriented.

Notable additions:

model warm pools and resident model limits

dynamic REST batching

request coalescing for identical in-flight work

rate-limit and concurrency-limit middleware

split /livez and /readyz probes with graceful shutdown behavior

trusted-host and CORS configuration

optional Prometheus-style metrics without PHI-derived labels

Docker Compose with HF cache volume

Kafka streaming de-identification connector

typed Python REST client

TypeScript REST client

Swift OpenMedKit de-identification JSON export and bundled policy profiles

Evaluation, Risk, And Release Evidence

The evaluation system now produces more granular evidence for privacy and model quality.

New capabilities include:

DrugProt public relation-extraction evaluation

biomedical NER benchmark suite

i2b2 eval-only loader

clinical PHI dataset manifest

dataset cards without row text

fixture coverage reports

per-section recall reports

eval result caching

model fleet freshness metrics

fairness, robustness, and error-analysis reports

leakage heatmaps by label and language

membership-inference and linkage attack modes

k-anonymity, l-diversity, and t-closeness reports

audit-report diffs

model scorecards

threshold sweeps and paired significance testing

flaky-run variance detection

calibration reliability data

over-redaction and utility-loss reports

policy-profile compliance evaluation

cross-release benchmark history diffs

risk dashboard rendering

release-gate dry-run previews and evidence bundles

Models, Backends, And Training

v1.7.0 adds model export and training infrastructure across Apple, browser, PyTorch, and quantized paths.

New work includes:

Laneformer MLX language model runtime

MLX INT4 recall certification

Core ML INT8 palettized export

AWQ and GPTQ 4-bit recipes

bitsandbytes 4-bit load-time quantization

FlashAttention, SDPA, and eager attention backend selection

PyTorch Metal/MPS device selection and tuning

ONNX/WebGPU export artifacts

Transformers.js browser export target

tokenizer caching

Mode-A knowledge distillation

DAPT corpus assembler

DirectID tiny-head contract

hard-negative training sampler

span-relation graph decoder

ONNX and quantized artifact publishing metadata

Security, Supply Chain, And Docs

The release also hardens project operations.

Additions include:

root SECURITY.md and private vulnerability reporting guidance

contributor onboarding and community health files

breach notification runbook and breach report template

CycloneDX SBOM generation and release attachment

reproducible-lock workflow

lockfile drift gate

GitHub Actions ref validation

manifest schema validation command and CI gate

release docs for trust status, clinical context, SBOM, reproducible dependencies, OpenAPI, REST clients, quantization, and policy workflows

doctest-backed examples for public APIs

explicit UTF-8 file/subprocess handling, subprocess timeouts, and lazy logging formatting fixes

Fixes

Audit span records now sort deterministically across runs.

Date shifting avoids zero-day offsets, preserves shifted intervals by default, and handles slash/dash two-digit-year formats more consistently.

Unicode offset redaction in the pipeline was fixed.

Duplicate FHIR resource IDs are rejected.

JSON loading paths in core, eval, NER, and risk modules raise clearer errors or fail closed on corrupt JSON.

Optional-extra failures now produce clearer skipped-capability metadata.

Safety sweeps reduce over-redaction from bare numeric identifier patterns.

GitHub Actions ref validation and setup-uv pinning were tightened.

Manifest validation now runs through uv.

The CLI fallback message for the basic TUI entry point was corrected.

Migration Notes

analyze_text(..., output_format="dict") now returns a frozen AnalyzeResult; to_dict() and mapping access preserve the legacy payload shape.

Clinical term protection is enabled by default in PII extraction and the staged pipeline, which can suppress ambiguous PERSON, LOCATION, or ORG matches that exactly match protected clinical vocabulary.

FHIR OperationOutcome output emits R4 issue.expression; legacy issue.location is accepted on input but is not emitted.

ServiceRuntime.get_loader() now returns the warm-pool proxy. Use get_model_loader() when raw loader access is required.

Custom OCR engines should accept the keyword-only languages parameter.

OCR auto-selection can choose optional EasyOCR or docTR engines when they are installed.

REST deployments using custom Host headers must configure OPENMED_SERVICE_TRUSTED_HOSTS; wildcard CORS and trusted-host settings are rejected.

Unsupported Core ML architectures now fail before model loading/tracing, and --quantized-output requires --quantize int8.

The canonical label set expanded with clinical concept labels.

format_preserve expands the de-identification action enum and schema surface.

The legacy shift_dates boolean remains accepted, but new code should prefer method="shift_dates" with explicit date-shift options.

PR Review Summary

Reviewed range: v1.6.0..origin/release/openmed-170

Last release: v1.6.0 at 1863350b

Reviewed head: origin/release/openmed-170 at 36ae8666

First-parent commits reviewed: 189

Merged PR commits reviewed: 148

Additional squash/direct PRs covered by these notes: 36

Total PRs covered by these notes: 184

Direct maintenance commits without a new PR number reviewed: e7f080b4 (revert grouped Actions update), 38e86172 (pin setup-uv action), cd3f2b9e (run manifest validation through uv), d1061d7b (README update), and 79d56791 (avoid MkDocs autorefs in doctest outputs)

Aggregate diff: 483 files changed, 87,085 insertions, 1,222 deletions

Major PR groups:

Multimodal, OCR, and document redaction: #555, #567, #717, #749, #558, #726, #745, #755, #758

Interop and structured health data: #566, #642, #631, #629, #626, #625, #553, #557, #737, #777, #784, #689, #690, #704, #372

Clinical extraction and context: #552, #410, #560, #568, #565, #718, #683, #773, #684, #691, #738, #739, #782, #785, #774, #698

Language, locale, and identifiers: #562, #747, #746, #748, #709, #609, #610, #614, #766

Evaluation, risk, and release gates: #617, #615, #743, #701, #688, #703, #702, #725, #680, #724, #723, #740, #735, #681, #682, #753, #754, #762, #765, #734, #764, #744, #786, #337, #351, #633, #634, #635, #636, #637, #638, #732, #379

Service, CLI, and SDKs: #632, #630, #750, #742, #722, #788, #789, #756, #741, #721, #780, #771, #772, #730, #775, #787, #696, #728, #776, #564

Models, backends, and training: #644, #620, #619, #627, #759, #760, #761, #719, #736, #790, #383, #751, #622, #612, #349, #603, #1020

Security, docs, CI, and release engineering: #648, #647, #716, #720, #1083, #693, #700, #607, #710, #711, #712, #713, #714, #715, #1021, #409, #697, #763, #185, #640, #646, #604, #605, #407

Open Items Before Tagging

Run the full project lint, test, build, and release-gate suite if you want coverage beyond the focused release-prep gates listed below.

Confirm the v1.7.0 release-gate candidate bundle has required metadata, calibration evidence, span fixtures, quantization evidence, SBOM, and reproducible-lock evidence.

If the GitHub release body is generated automatically, confirm it is based on v1.6.0...v1.7.0 and keep the Fixes #399 issue reference out of the PR count.

Decide whether CHANGELOG.md should be expanded to include the 36 squash/direct PRs as explicit PR-number references, or whether the detailed release-note inventory is the source of truth for that level of traceability.

Verification For This Draft

The v1.6.0 GitHub release body and local RELEASE_NOTES_v1.6.0.md were used as the style reference.

The reviewed range is v1.6.0..origin/release/openmed-170; current reviewed head is 36ae8666.

The first-parent range contains 189 commits: 148 merge commits plus 41 direct first-parent commits.

GitHub generated release notes for v1.7.0 resolve to 184 PR links. The extra #399 in a commit title is an issue reference, not a PR, and is not counted as a PR.

Remaining 1.6 search hits after the version bump are historical changelog entries, compare ranges, CycloneDX/spec/style/dependency values, or changelog-generator examples.

python3 scripts/release/check_release_version.py --version 1.7.0 passed.

python3 scripts/release/check_repo_policy.py passed.

Focused tests passed: 70 passed, 2 warnings for release changelog, FHIR provenance, audit/risk CLI, OpenAPI spec, and service API tests.

make docs-build passed under MkDocs strict mode.

git diff --check passed.

What's Changed

feat: declare the [multimodal] extra and ingest/redact contract by @pardeep-singh in #555

feat:added FHIR OperationOutcome builder for exporter and validation results by @Sidharth1743 in #566

feat: add dermatology and ophthalmology domains to NER catalog by @pardeep-singh in #552

feat: add Tesseract and PaddleOCR engine adapters with a common OCR result contract by @pardeep-singh in #567

Build the hard-negative harness for training batches by @maziyarpanahi in #349

Feature/lab parser by @VishnuPrasath-S-20 in #568

Add per-language locale and national-ID surrogate coherence regression suite by @cycsmail in #562

Add the LangChain redaction wrapper integration spike by @maziyarpanahi in #373

Define the nightly trust results schema by @maziyarpanahi in #379

Add ONNX and WebGPU export artifacts by @maziyarpanahi in #383

Define DirectID tiny head contract by @maziyarpanahi in #603

Add clinical PHI flagship dataset manifest by @maziyarpanahi in #388

fix: correct misleading message in _run_basic_tui_entry fallback (Fixes #399) by @rtmalikian in #407

fix: reject duplicate resource ids in FHIR Bundle assembler by @gitcommit90 in
Note truncated.

Additional notes

This release summarizes 148 pull requests merged into release/openmed-170 after v1.6.0. The diff is additive overall: 483 files changed, with no deleted or renamed files detected in the release range.

Added

Added lightweight multimodal document primitives, source spans, lazy handler registration, redact_document, image redaction, PDF span coordinate projection, Markdown/AsciiDoc offset-preserving extraction, audit-safe image, PDF, and DOCX metadata scrubbing, and JSONL chat-log de-identification with speaker pseudonymization (#555, #567, #726, #745, #755, #758).
Added OCR engine coverage for Tesseract, PaddleOCR, EasyOCR, docTR, and test engines, including OCR language selection and available-engine discovery (#567, #717, #749, #558).
Added CDA/C-CDA XML, HL7 v2, CSV/TSV, FHIR $de-identify, FHIR Bulk NDJSON, deterministic FHIR Bundle, FHIR OperationOutcome, FHIR Provenance / AuditEvent, deterministic urn:uuid, code-system provenance, CodeableConcept checks, and flat-table clinical entity export helpers (#566, #642, #631, #629, #626, #625, #553, #705, #737, #777, #784, #689, #690).
Added clinical extraction and normalization helpers for labs, vital signs, medication sigs, problem lists, summary cards, microbiology labels, dermatology and ophthalmology domains, clinical concept labels, and clinical term protection, plus deterministic substance, employment, and living-status normalization (#552, #410, #560, #718, #683, #773, #684, #691, #698, #767).
Added a nutrition and diet-order zero-shot domain, four canonical nutrition policy labels, policy-profile coverage, routing metadata, and synthetic fixture coverage for diet orders and feeding routes (#951).
Added language and locale capabilities for Indonesian, Thai, Hebrew RTL, PESEL, Korean RRN, Unicode script detection, locale checksum registries, deterministic locale PHI generation, and locale-aware date/number normalization (#747, #746, #748, #709, #609, #610, #614, #766).
Added de-identification runtime features: DeidentificationResult.to_dataframe, redaction preview diffs, cross-document surrogate vaults, patient-keyed date shifting, format-preserving identifier redaction, minimum-necessary strength selection, streaming incremental de-identification, typed analyze results, pipeline explain traces, section stamping, and per-document risk budgets (#706, #695, #729, #704, #778, #779, #731, #611, #727, #785, #733).
Added CLI surfaces for policy-aware openmed deid, openmed fhir bundle, openmed models recommend, openmed models diff, openmed policy diff, openmed doctor, openmed gates preview, openmed gates bundle, openmed audit, openmed risk, and active-learning queue management (#741, #777, #721, #780, #771, #772, #775, #735, #787, #613).
Added service features for model warm pools, dynamic batching, request coalescing, rate and concurrency limits, readiness/liveness endpoints, opt-in Prometheus metrics, and typed Python and TypeScript REST clients (#632, #630, #750, #742, #722, #788, #789, #756).
Added an in-process ASGI load-test harness with configurable concurrency that reports requests per second, p50/p95/p99 latency, and error rate (#461).
Added evaluation, release-gate, and risk tooling: DrugProt and public biomedical NER suites, i2b2 loader, multilingual golden fixtures, dataset cards, fixture coverage, per-section recall, result cache, leakage heatmaps, membership-inference probe, k-anonymity/l-diversity/t-closeness metrics, audit diffs, evidence bundles, scorecards, threshold sweeps, flaky-run detection, paired significance testing, calibration reliability data, utility-loss reports, policy-compliance suite, cross-release benchmark history diffs, nano-tier certification, and risk dashboard rendering (#617, #615, #743, #701, #688, #703, #702, #708, #725, #680, #724, #723, #740, #735, #681, #682, #752, #753, #754, #762, #765, #734, #764, #744, #786).
Added model, backend, and training support for Laneformer MLX-LM, MLX INT4 recall certification, Core ML INT8 palettized export, AWQ and GPTQ 4-bit quantization recipes, bitsandbytes 4-bit loading, FlashAttention/SDPA/eager attention selection, PyTorch MPS tuning, ONNX/WebGPU and Transformers.js exports, tokenizer caching, Mode-A distillation, DAPT corpus assembly, and ONNX/quantized artifact publishing metadata (#644, #620, #619, #627, #759, #760, #761, #719, #736, #790, #751, #622, #612).
Added interop adapters for PHILTER, pyDeid, GLiNER-BioMed, LangChain, and the optional spaCy openmed_deid pipeline component (#372, #624).
Added policy profiles and policy tooling for Australia Privacy Act, GDPR Article 9 health, UK ICO anonymisation, policy config diffing, and Swift OpenMedKit policy-driven de-identification (#769, #770, #768, #771, #685).
Added Swift/OpenMedKit de-identification result JSON export and bundled policy resources for client-side policy workflows (#692, #685).
Added examples and documentation for a first-five-minutes redaction/extraction to FHIR walkthrough, OpenAPI export, model manifest docs, REST clients, OCR, multimodal redaction, quantization exports, policy workflows, security, SBOM, reproducible dependencies, breach response, onboarding, community health, and release status contracts (#628, #694, #647, #716, #720, #1021, #409, #697).

Changed

analyze_text(..., output_format="dict") now returns a frozen AnalyzeResult; to_dict() and mapping access preserve the legacy dict shape (#611).
PII extraction and the staged pipeline now apply clinical term protection by default, suppressing ambiguous PERSON/LOCATION/ORG matches that exactly match protected clinical vocabulary (#698).
ConText temporality, uncertainty, and negation now use sentence/clause-bounded cue scope with section-aware priors and context offsets (#738, #739, #782).
Pipeline span output can include populated section metadata after section stamping (#785).
Lab reference-range parsing now accepts broader separators/operators and treats unknown explicit flags as unknown rather than deriving a normal/high/low result (#560).
REST /health remains as a compatibility alias, while /livez and /readyz expose split liveness/readiness state and shutdown drains in-flight model-backed requests (#722).
REST CORS and trusted-host handling is now deny-by-default except for exact configured origins and trusted hosts (#686).
OCR auto-selection can now pick installed EasyOCR or docTR adapters in addition to Tesseract/PaddleOCR (#749, #558).
Evaluation defaults now include DrugProt and biomedical NER suites, and leakage heatmaps now emit label-by-language matrices with totals and worst cells (#617, #743, #680).
Model manifest rows now merge format lists for existing repositories and recognize ONNX/WebGPU and Transformers.js export formats (#736, #790).
CI lint/test/security/build setup moved to uv sync / uv run, with GitHub Actions refs validated and Dependabot Actions updates limited to minor/patch bumps (#185, #700).
PyTorch/HF backends can auto-select MPS on Apple Silicon when no device is set (#719).
AWQ and GPTQ export paths now share synthetic quantization calibration metadata (#759).
shift_dates documentation now describes patient-keyed stable date shifting; the legacy boolean remains accepted but deprecated in favor of method="shift_dates" (#704).

Fixed

Fixed nondeterministic audit span ordering so report serialization, hashes, and signatures are stable while preserving legacy verification (#645).
Fixed date-shift parity between python-dateutil and fallback paths, including month-first English month-name dates, and aligned uv.lock with the dev extra dependency set (#616, #649).
Fixed deterministic FHIR URN preservation during Bundle assembly (#553).
Fixed JSON loading paths in core, eval, NER, and risk modules so corrupt JSON raises clearer errors or fails closed (#958).
Fixed optional-extra diagnostics for missing ftfy, section detection, and date-shift capabilities (#781).
Reduced numeric false positives in safety-sweep postcode-style matches by requiring stronger context (#783).
Added explicit UTF-8 encodings for subprocess/file I/O paths and preserved exception chaining in model load failures (#1088).
Added timeouts to subprocess.run calls in reproducibility hash and release-gate issue helpers (#1090).
Replaced eager f-string logging with lazy logging interpolation across model, processing, batch, text, and utility modules (#1092).
Fixed PII method quickstart docs for mask, remove, replace, hash, shift_dates, and reidentify() examples (#409).

Security

Added root SECURITY.md, private vulnerability disclosure guidance, security issue-template routing, security docs, and README links (#648).
Added breach-notification runbook and breach report template with explicit no-raw-PHI/PII handling guidance (#1021).
Added CycloneDX SBOM generation via make sbom, CI artifact upload, tagged release SBOM attachment, and supply-chain docs (#720).
Added reproducible-lock GitHub Actions gate and contributing docs for pinned, hash-verified installs (#1083).
Added lockfile drift, GitHub Actions ref, license-policy, and doctest-backed public-example gates (#693, #700, #763).
Added PHI-safe defaults for progress callbacks, NDJSON error summaries, active-learning records, hashed examples, explain traces, dataset cards, and metadata scrubbing (#621, #737, #613, #765, #727, #701, #755).

Dependencies

Added optional extras and dependency policy entries for multimodal/OCR, spaCy, AWQ, GPTQ, MLX-LM, Kafka, PHILTER/pyDeid, TypeScript client support, and service clients (#555, #567, #624, #627, #644, #757, #759, #372, #756, #789).
Updated GitHub Actions refs and maintenance dependencies, including checkout v7, setup-python v6, cache v6, upload-artifact v7, Ruff/pre-commit updates, and LangChain Core 1.x compatibility for the optional LangChain extra (#607, #710, #711, #712, #713, #714, #715).

Removed

No public files, modules, or APIs were removed in the reviewed release range.

Upgrade Notes

FHIR OperationOutcome output emits R4 issue.expression; legacy issue.location is accepted on input but is not emitted, and non-R4 severities such as info are rejected (#566).
ServiceRuntime.get_loader() returns the warm-pool proxy; use get_model_loader() when raw loader access is required (#632).
Unsupported Core ML architectures now fail before model loading/tracing, and --quantized-output requires --quantize int8 (#619).
Custom OCR engines should tolerate the keyword-only languages parameter (#717).
The canonical label set expanded with clinical concepts, which can affect callers enumerating exact label counts (#718).
format_preserve expands the action enum/schema surface and updates schema fingerprints (#778).
REST deployments using custom Host headers must configure OPENMED_SERVICE_TRUSTED_HOSTS; wildcard CORS/trusted-host settings are rejected (#686).
OCR auto-selection order changed when optional EasyOCR or docTR engines are installed (#749, #558).

Open source →

1.6.023 Jun 2026

OpenMed 1.6.0 is the privacy assurance and release-governance release.

OpenMed `1.6.0` is the privacy assurance and release-governance release.

This release turns de-identification from a single API call into a full, auditable privacy system: policy-aware pipelines, canonical span records, deterministic safety backstops, calibrated thresholds, signed audit reports, re-identification risk scoring, and release gates that fail closed when evidence is missing.

It also adds the infrastructure needed to ship model artifacts responsibly: manifest-driven model cards, Hugging Face publishing controls, dependency and secret-scanning gates, public benchmark reporting, and a richer evaluation suite.

Highlights

Added a policy-aware de-identification runtime built around canonical OpenMedSpan records, a ten-stage privacy pipeline, detector arbitration, calibrated thresholds, deterministic safety sweeps, and six bundled policy profiles.
Added signed, reproducible audit reports with span provenance, residual-risk metadata, reproducibility hashes, and optional HMAC signatures.
Added re-identification risk reporting and adversarial re-identification benchmark mode.
Added a leakage-first evaluation harness with synthetic golden fixtures, public/reference dataset adapters, DUA-gated stubs, SHIELD support, weak labeling, cold-start latency, and bootstrap confidence intervals.
Added v1.6.0 release-gate infrastructure: last-green baselines, calibration artifacts, G1a-G8 signed reports, quantization recall-delta checks, status/leaderboard generation, and fail-closed CI.
Added clinical and interoperability utilities: ConText temporality and uncertainty axes, OHDSI Athena/Usagi ingestion, a Presidio bridge, and deterministic FHIR R4 transaction/batch Bundle assembly.
Added a cardiology zero-shot label-map domain plus cardiology keyword routing metadata for future model registration.
Added a canonical models.jsonl manifest, manifest refresh workflow, manifest-driven Hugging Face model cards, and HF publishing for converted MLX/CoreML artifacts.
Added a packaged openmed CLI surface with benchmark and calibration commands, plus a de-identification cookbook notebook and clinical NER families example.
Added governance, compliance, security, device-tier, FAQ, API reference, release-channel, status, leaderboard, and notebook documentation.

The Big Picture

OpenMed v1.6.0 is about moving from "the model found PII" to "the system can prove what happened."

The new privacy stack records normalized spans, detector provenance, policy actions, thresholds, safety-sweep findings, risk metadata, and stable hashes. That makes de-identification outputs easier to audit, reproduce, benchmark, and review before they reach clinical, operational, or demo workflows.

The same theme shows up in release engineering. Model and package releases now have stronger evidence requirements: calibrated thresholds, last-green baselines, quantization deltas, release gate reports, dependency/license checks, and secret scanning.

Privacy Pipeline

deidentify() now routes through a staged privacy pipeline.

The pipeline includes:

Unicode normalization and offset mapping
language/script routing
section and document-type metadata
deterministic PII detectors
fast PII and clinical PHI model hooks
detector arbitration and cascade routing
policy action assignment
deterministic safety sweep
auditable result emission

Bundled policy profiles:

hipaa_safe_harbor
hipaa_expert_review_assist
gdpr_pseudonymization
research_limited_dataset
strict_no_leak
clinical_minimal_redaction

Example:

from openmed import deidentify

result = deidentify(
    "Patient Jane Rivera, MRN 123456, was seen on 2026-06-01.",
    method="replace",
    policy="hipaa_safe_harbor",
    keep_mapping=True,
    audit=False,
)

print(result.deidentified_text)

Use audit=True when you need the signed audit-report path instead of the regular DeidentificationResult.

Audit And Risk

OpenMed now includes signed audit reports and re-identification risk scoring.

Audit reports can include:

canonical span records
detector and policy provenance
threshold metadata
residual-risk summaries
reproducibility hashes
optional HMAC signatures

Risk reporting adds leakage and re-identification views for text and tabular records, including singleton and quasi-identifier analysis. The new adversarial re-identification benchmark mode lets release candidates be tested against round-trip and surrogate-consistency failures.

Evaluation And Release Gates

The new evaluation framework is leakage-first. It adds BenchmarkReport, golden fixtures, SHIELD comparison support, weak labeling, public/gated dataset adapters, cold-start latency, bootstrap confidence intervals, and report rendering.

Release gates now cover G1a-G8 readiness checks with signed reports and fail-closed CI. The release-gate workflow expects evidence for:

release metadata
calibration thresholds
masking and replacement profiles
G8 span fixtures
quantization recall deltas for quantized formats
last-green baseline comparisons

This is not only for dashboards. It is meant to prevent a model artifact from being promoted without the evidence needed to defend the release.

Clinical And Interoperability

v1.6.0 adds several clinical interoperability building blocks:

cardiology zero-shot labels: CardiacFinding, ECGFinding, EjectionFraction, CardiacProcedure, CardiacDevice, and Anatomy
cardiology keyword routing metadata for ECG findings, ejection fraction, arrhythmias, stents, pacemakers, murmurs, echocardiograms, and related cardiology text
ConText temporality axis: recent, historical, hypothetical
ConText uncertainty axis
OHDSI Athena vocabulary loader
Usagi mapping ingestion
Presidio adapter
FHIR R4 Bundle assembler

The FHIR assembler produces deterministic transaction/batch Bundles with stable urn:uuid fullUrl values, internal reference rewriting, and request metadata for server ingestion.

No cardiology model is registered in v1.6.0. Cardiology routing is forward metadata and zero-shot label-map support; public model suggestions still fall back to existing general medical recommendations until a cardiology model ships.

Model Manifest And Publishing

The canonical model manifest is now a first-class artifact.

This release adds:

models.jsonl
manifest refresh workflow
manifest-driven model registry loading
manifest-driven model card generation
Hugging Face upload support for converted MLX/CoreML artifacts
protected hf-publish environment policy
HF_WRITE_TOKEN handling guidance

Packaging now includes the manifest, gate baseline, policy/schema JSON, LICENSE, and NOTICE so release artifacts have the metadata they need.

CLI, Docs, And Examples

The openmed CLI is now packaged and includes benchmark and calibration commands. The benchmark command supports SHIELD suite runs, multi-model JSON output, and adversarial re-identification mode.

New and refreshed documentation includes:

de-identification cookbook notebook
clinical NER families example
project FAQ
mkdocstrings API reference
compliance posture
release streams and channels
HF token policy
dependency and secret-handling docs
device-tier SLOs
status and leaderboard pages

Security And CI

v1.6.0 expands release hardening:

protected HF publish environment
dependency license policy gate
pip-audit gate with time-boxed ignores
gitleaks CI/pre-commit secret scanning
grouped GitHub Actions Dependabot updates
repo policy checks
Ruff-based Python lint/format
Swift swift-format scripts and CI gate
release-gates workflow

The old duplicate .github/workflows/release.yml workflow was removed. PyPI publishing now goes through the guarded tag/manual publish.yml workflow.

Fixes

REST/MCP PII schemas now accept ar, ja, and tr.
Privacy-filter trust_remote_code allowlist matching is case-insensitive.
Feb 29 date shifting clamps safely when keep_year=True targets a non-leap year.
REST requests are capped by OPENMED_SERVICE_MAX_TEXT_LENGTH and return 422 on oversized text.
BatchProcessor.iter_process now honors batch_size while preserving order.
Duplicate benchmark fixture IDs are rejected.
Duplicate benchmark CLI registration was fixed.
Release gates skip cleanly when no candidate report exists.
User-controlled HTML formatter fields are escaped.
Long non-ASCII/CJK clinical text is no longer rejected as suspicious content.
Reversible remove mappings and repeated same-type placeholders now work with keep_mapping=True.

Migration Notes

deidentify(..., audit=True) returns an audit report. Leave audit=False when callers expect DeidentificationResult.
deidentify(..., keep_mapping=True) can now produce unique repeated placeholders such as [NAME_2]. Snapshot tests may need updates. Without keep_mapping, legacy repeated mask placeholders are preserved.
Prefer method="shift_dates" in new code. shift_dates remains available as a compatibility alias.
REST clients sending more than OPENMED_SERVICE_MAX_TEXT_LENGTH characters now receive a 422 unless the limit is raised.
Full SHIELD and DUA corpora require approved or user-supplied access paths; restricted rows are not vendored.
Use .github/workflows/publish.yml for PyPI release operations.
Model publishing requires HF_WRITE_TOKEN in the protected hf-publish environment.
Local Python formatting is now Ruff-based: make format, make lint, and make format-check.
Swift formatting uses make format-swift and make lint-swift.

PR Review Summary

Reviewed range: v1.5.5..master

Last release: v1.5.5 at adff5fe
Reviewed head: origin/master at 3b62f2e
Merged PRs reviewed: 69
Direct first-parent release-prep commits reviewed: 7 (f654531, .gitignore; 8f398cc, CHANGELOG.md; baad551, CHANGELOG.md; 0769007, docs/swift-openmedkit.md; 0190625, docs/website/index.html; 0b9bdd8, openmed/__about__.py; 3b62f2e, README.md)
Aggregate diff: 371 files changed, 40,102 insertions, 4,854 deletions

Major PR groups:

Privacy/de-identification: #193, #194, #198, #208, #210, #218, #323, #341, #342, #343, #344, #345, #346, #371, #559
Evaluation/release gates: #195, #203, #207, #338, #339, #353, #354, #357, #358, #360, #361, #369
Clinical/interop/zero-shot: #332, #334, #347, #368, #370, #550, #551, #556
Release/security/publishing: #186, #220, #221, #326, #331, #333, #364, #366, #367, #412, #554
Docs/examples/website/CLI/service: #65, #187, #196, #197, #200, #201, #206, #212, #214, #219, #322, #355, #359, #362, #390, #401, #403, #404, #406
Training/datasets/model registry: #188, #191, #348, #350

Open Items Before Tagging

Reconcile the website language-support copy: PR #404 advertises 16-language PII support, while core/service/README/FAQ currently enumerate 12 PII languages (en, fr, de, it, es, nl, hi, te, pt, ar, ja, tr).
Confirm the v1.6.0 release-gate candidate bundle has required metadata, calibration evidence, G8 span fixtures, and quantization evidence.
Previously identified release blockers are now resolved by merged PRs: #554 / #529 rejects empty audit HMAC keys; #550 / #530 rejects duplicate FHIR ResourceType/id values; #559 / #513 shifts default-model lowercase date labels by canonical label.
Decide whether #68 is required by release policy. GitPython is present in uv.lock through the docs extra, not imported by runtime code.
#556 was merged and adds FHIR Bundle empty-input and dangling-reference regression coverage. Treat #553 as a useful non-blocking FHIR refactor unless its helper extraction is required for the release branch.
Run the full project test/lint/release-gate suite before tagging if you want coverage beyond the targeted release-prep gates listed below.

Verification For This Draft

Re-audited against current origin/master on 2026-06-22 after PRs #559, #554, #550, and #556 plus the direct release-prep commits through 3b62f2e were merged. The latest release tag is still v1.5.5, reviewed head is 3b62f2e, and the first-parent range contains 76 commits: 69 merged PRs plus direct release-prep commits f654531, 8f398cc, baad551, 0769007, 0190625, 0b9bdd8, and 3b62f2e.
Four parallel subsystem reviews were completed: privacy/de-identification, evaluation/release gates, docs/CLI/service, and CI/release/MLX.
The pasted GitHub release-note list, live gh pr list metadata, and this draft resolved to the same 64 merged PR numbers before PR #551. After PR #551, #559, #554, #550, and #556, live GitHub metadata and this draft resolve to the same 69 merged PR numbers, with no missing or extra PRs.
A changed-area coverage check confirmed explicit release-note coverage for the privacy pipeline, evaluation harness, release gates, datasets, clinical interop, model manifest/publishing, CLI/service, docs/examples, website, security/CI, Swift/MLX, tests, packaging, and the direct .gitignore commit.
PR #551 was reviewed separately: CI passed, focused cardiology label/routing tests passed locally, Ruff passed on changed files, and the runtime probe confirmed cardiology zero-shot labels plus private cardiology routing while public model suggestions retain the intended fallback behavior.
PRs #559, #554, #550, and #556 were reviewed separately with parallel focused checks. method="shift_dates" now uses canonical date labels, de-identification audit signing rejects empty HMAC keys, FHIR Bundle assembly rejects duplicate ResourceType/id values, and empty/dangling FHIR Bundle edge cases now have regression coverage.
Latest origin/master CI for 3b62f2e passed.
python3 scripts/release/check_release_version.py --version 1.6.0 passed after the local version bump.
uv run --extra docs mkdocs build --strict passed.
uv run --extra dev ruff check . passed.
uv run --extra dev pytest tests/unit -q passed: 1,615 passed, 1 skipped.
uv run --with build python -m build passed and produced 1.6.0 artifacts.
uv run --with twine twine check dist/* passed.
git diff --check passed after the changelog, version, and release-note edits.

What's Changed

docs: on-device iOS showcase + MLX/batch performance visuals by @maziyarpanahi in #65
Generate canonical model manifest by @maziyarpanahi in #191
Scaffold section 4.2 package shells by @maziyarpanahi in #188
Track CLI and MCP packages in git by @maziyarpanahi in #187
Scaffold the openmed/eval benchmark harness with leakage-first metrics by @maziyarpanahi in #195
Website: responsive mobile nav, 6M+ downloads stat, contribute CTA by @maziyarpanahi in #206
Add the deterministic post-ML safety sweep recall backstop by @maziyarpanahi in #193
Add re-identification risk scoring (risk_report) for text and tables by @maziyarpanahi in #194
Commit synthetic golden de-identification fixtures for the eval suites by @maziyarpanahi in #203
Add 6M+ PyPI downloads stat and contribute call-to-action by @maziyarpanahi in #212
Add favicon, touch icon, and logo.svg for the website by @maziyarpanahi in #214
Fix case-sensitive trust_remote_code allowlist matching by @sawirricardo in #208
fix: accept ar/ja/tr in REST/MCP PII lang schema by @guhyun9454 in #210
docs: complete Hindi README to parity with README.md by @Technolity in #196
docs: fix non-functional analyze_text example in getting-started by @Technolity in #197
fix: clamp Feb 29 when keep_year date shifting hits a leap day by @Technolity in #198
fix: cap request text length in the REST service by @Technolity in #200
fix: respect batch_size in BatchProcessor.iter_process by @Technolity in #201
Consolidate PyPI publish into a single tag-triggered release workflow by @maziyarpanahi in #186
Provision write token and CI rotation policy by @maziyarpanahi in #220
Add HF push_to_hub publishing to the model conversion pipeline by @maziyarpanahi in #333
Add the last-green baseline store and reproducibility hash by @maziyarpanahi in #338
Extend label policy metadata and HIPAA cross-map by @maziyarpanahi in #323
Define the canonical OpenMedSpan record and ship versioned JSON schemas by @maziyarpanahi in #342
Build the staged privacy detection pipeline runtime by @maziyarpanahi in #343
Implement detector arbitration with the cascade router and score calibration by @maziyarpanahi in #344
Add the versioned per-label/language/policy thresholds matrix by @maziyarpanahi in #345
Implement the six policy profiles and the deidentify(policy=) and REST selector by @maziyarpanahi in #346
Author the standardized training recipe package with three presets by @maziyarpanahi in #348
Add cold_start_ms measurement to the eval harness by @handlecusion in #353
Add project FAQ page by @Vishnupriyaaaaaaaaaaaa in #322
add clinical NER families example by @alberthammerich in #355
Build signed reproducible audit report with span provenance by @maziyarpanahi in #326
Promote quality_gates span overlap detection to deterministic resolution by @maziyarpanahi in #218
Fit per-(model,label,language) calibration thresholds by @maziyarpanahi in #357
Implement the ReleaseGate harness scoring G1a-G8 with a signed gate report by @maziyarpanahi in #360
Add mkdocstrings API reference generation by @yuvrajsingh2428 in #362
Add dependency audit gate by @alberthammerich in #364
Add a CI secret-scanning gate to block committed credentials by @maziyarpanahi in #367
build(deps): bump the github-actions group with 10 updates by @dependabot[bot] in #366
reject duplicate fixture IDs by @alberthammerich in #354
feat: generate model cards from the manifest at publish time by @pardeep-singh in

Note truncated.

Open source →

Additional notes

Added

Added a policy-aware de-identification runtime with canonical OpenMedSpan schema contracts, a ten-stage Pipeline, detector arbitration/cascade routing, calibrated per-label/language/policy thresholds, deterministic safety sweep backstops, and six bundled policy profiles (hipaa_safe_harbor, hipaa_expert_review_assist, gdpr_pseudonymization, research_limited_dataset, strict_no_leak, clinical_minimal_redaction).
Added signed, reproducible de-identification audit reports with span provenance, residual-risk metadata, reproducibility hashes, and optional HMAC signatures.
Added re-identification risk reporting and adversarial re-identification benchmark support, including openmed benchmark pii --attack reid.
Added a leakage-first evaluation harness with BenchmarkReport, synthetic golden de-identification fixtures, public/reference dataset adapters, DUA-gated corpus stubs, SHIELD comparison-suite support, weak labeling utilities, cold-start latency, and deterministic bootstrap confidence intervals.
Added release-gate infrastructure for v1.6.0 model readiness: last-green baselines, calibration artifacts, G1a-G8 signed gate reports, quantization recall-delta checks, generated status/leaderboard pages, and a fail-closed release-gates workflow.
Added clinical and interoperability utilities: ConText temporality and uncertainty axes, OHDSI Athena/Usagi ingestion, a Presidio adapter, and a deterministic FHIR R4 transaction/batch Bundle assembler.
Added a cardiology zero-shot label-map domain (CardiacFinding, ECGFinding, EjectionFraction, CardiacProcedure, CardiacDevice, Anatomy) plus cardiology keyword routing metadata for future model registration. Public model suggestions continue to fall back to existing general medical models until a cardiology model is registered.
Added a canonical models.jsonl manifest, manifest refresh workflow, manifest-driven Hugging Face model card generation, and HF publishing support for converted MLX/CoreML artifacts.
Added a packaged openmed CLI surface with benchmark and calibration commands, plus a de-identification cookbook notebook and an offline clinical NER families example.
Added governance, compliance, security, device-tier, FAQ, API reference, release-channel, status, leaderboard, and notebook documentation.

Changed

deidentify() now routes through the staged policy pipeline and accepts policy, calibration, threshold, and audit controls. When audit=True, it returns an audit report rather than the regular DeidentificationResult.
deidentify(..., keep_mapping=True) now emits unique placeholders for repeated entities of the same type, such as [NAME] and [NAME_2], so re-identification round trips can distinguish them.
Label metadata now carries policy labels, HIPAA Safe Harbor mappings, risk levels, and ID-number subtype hints while keeping canonical labels stable.
Benchmark steady-state latency now excludes cold start while preserving latency.cold_start_ms in reports.
PyPI publishing now uses a single guarded tag/manual publish.yml workflow; the duplicate release workflow was removed.
Release metadata now derives changelog sections and expected SemVer bumps from Conventional Commits.
Python linting/formatting moved to Ruff and pre-commit, Swift formatting moved to checked-in swift-format scripts, and CI now enforces the updated repo policy, lint, tests, security, secret-scan, Swift-format, and release-gate jobs.
Packaging now includes the model manifest, release-gate baseline, policy/schema JSON, LICENSE, and NOTICE.

Fixed

Fixed method="shift_dates" to recognize canonical date labels before redaction, so lowercase date output from the default English PII model and date_of_birth labels are shifted instead of masked; keep_mapping no longer treats shifted dates as mask placeholders.
FHIR Bundle assembly now rejects duplicate ResourceType/id values instead of silently overwriting the earlier resource in the internal reference map. Duplicate resources raise a ValueError that names the colliding key, preventing downstream references from being rewritten to the wrong Bundle entry.
REST/MCP request schemas now accept ar, ja, and tr for the lang field. These languages have published PII models and are listed in SUPPORTED_LANGUAGES, but the lang Literal in openmed/service/schemas.py was never updated, so the service rejected them with a 422 even though the Python API and the models worked. The four lang annotations now share a single PIILanguage alias kept in sync with SUPPORTED_LANGUAGES (guarded by a regression test).
Fixed case-insensitive trust_remote_code allowlist matching for first-party and environment-configured privacy-filter repositories.
Fixed Feb 29 date shifting when keep_year=True targets a non-leap year.
Fixed REST oversized-text handling with OPENMED_SERVICE_MAX_TEXT_LENGTH (default 1_000_000 characters).
Fixed BatchProcessor.iter_process so batch_size is honored while preserving output order.
Fixed duplicate benchmark fixture IDs, duplicate benchmark CLI registration, release-gate behavior when no candidate report is present, and repo-policy ignored-file handling.
Fixed user-controlled HTML formatter escaping and validation false positives for legitimate long non-ASCII/CJK clinical text.
Fixed reversible remove mappings and repeated entity-type re-identification round trips when keep_mapping=True.

Security

Added a protected hf-publish environment and HF_WRITE_TOKEN policy for model publishing.
Added dependency license policy, pip-audit security gate with time-boxed ignores, and gitleaks CI/pre-commit secret scanning with a canary fixture.
Hardened de-identification audit report signing so AuditReport.sign() and AuditReport.verify() require a non-empty HMAC key. None, empty strings, and empty byte strings now raise ValueError instead of producing or accepting weak signatures.

Tests

Added FHIR Bundle regression coverage for empty resource lists across transaction, collection, and batch Bundles, and for dangling references that should remain unchanged when the referenced resource is absent from the Bundle.

Notes

shift_dates remains available as a compatibility alias; prefer method="shift_dates" in new code.
REST clients sending more than OPENMED_SERVICE_MAX_TEXT_LENGTH characters now receive a 422 response unless the limit is raised.
Full SHIELD/DUA datasets require approved or user-supplied access paths; restricted corpus rows are not vendored.
Release-gate candidates for v1.6.0 need release metadata, calibration evidence for masking/replacement profiles, span fixtures for G8, and quantization evidence for quantized formats.

1.5.58 Jun 2026

OpenMed 1.5.5 is a local-first PII, service lifecycle, and README refresh release.

OpenMed `1.5.5` is a local-first PII, service lifecycle, and README refresh release.

This release adds batch PII extraction/de-identification, REST model unload and keep-alive controls, stronger Swift/OpenMedKit long-document PII handling, and a polished multilingual README/brand refresh.

Highlights

Added BatchProcessor(operation="extract_pii") and BatchProcessor(operation="deidentify") for batch PII workflows.
Added REST model lifecycle controls: GET /models/loaded, POST /models/unload, request-level keep_alive, and OPENMED_SERVICE_KEEP_ALIVE.
Added Swift/OpenMedKit chunked PII extraction for long OCR text and improved ScanDemo clinical document PII recovery.
Refreshed the README with mascot/brand assets, an on-device PII demo GIF, Apple Silicon/Swift/iOS entry points, and 13 translated READMEs.
Bumped package, README, docs, website, and Swift demo version surfaces to 1.5.5.

What's Changed

add extract_pii/deidentify to BatchProcessor via an operation parameter by @thirdwing in #61
Add model unload and keep-alive controls by @maziyarpanahi in #62
[codex] Improve ScanDemo clinical document PII flow by @maziyarpanahi in #63
docs: refresh README with mascot, visuals & 14-language translations by @maziyarpanahi in #64

Full Changelog: v1.5.2...v1.5.5

What's Changed

add extract_pii/deidentify to BatchProcessor via an operation parameter by @thirdwing in #61
Add model unload and keep-alive controls by @maziyarpanahi in #62
Improve ScanDemo clinical document PII flow by @maziyarpanahi in #63
docs: refresh README with mascot, visuals & 14-language translations by @maziyarpanahi in #64

New Contributors

@thirdwing made their first contribution in #61

Full Changelog: v1.5.2...v1.5.5

Open source →

Additional notes

Added

Added batch PII extraction and de-identification support through BatchProcessor(operation="extract_pii") and BatchProcessor(operation="deidentify"), including document-level batch_size chunking, shared loader/pipeline reuse, tests, docs, and a runnable example.
Added REST service model lifecycle controls with GET /models/loaded, POST /models/unload, request-level keep_alive, OPENMED_SERVICE_KEEP_ALIVE, and model-loader cache release helpers.
Added chunked Swift/OpenMedKit PII extraction for long OCR text and refreshed the OpenMed Scan Demo clinical document flow with updated sample text, a printable sample PDF, and a generator script.
Added a project mascot, brand assets in docs/brand/, and an animated on-device PII de-identification demo (docs/brand/openmed-pii-demo.gif).
Added README translations in 13 languages with a language switcher: zh-CN, es, fr, de, it, pt, nl, ar, hi, te, ja, tr, fa.

Changed

Batched privacy-filter inference now accepts list inputs across Torch and MLX paths and forwards batching controls to the underlying pipelines.
The OpenMed Scan Demo now unloads inactive MLX runtime families when switching engines, sequences selected and secondary PII engine runs explicitly, improves OCR line ordering, and expands entity category mapping.
README and service/model-loader documentation now cover batch PII operations and model unloading behavior.
Overhauled the README with a visual hero, brand badges, Apple Silicon/Swift/iOS entry points, an OpenMed-vs-cloud comparison table, and a Mermaid flow diagram.

Fixed

Improved Swift structured PII recovery for clinical discharge summaries, including surname-first names, member and insurance IDs, account/encounter/document IDs, NPI values, PCP/signed-provider sections, and overlap deduplication.

Open source →

1.5.227 May 2026

This release fixes the privacy-filter remote-code trust boundary, tightens model-name routing for the OpenAI/OpenMed Privacy Filter family, and brings

OpenMed `1.5.2` is a security and MLX conversion hardening release.

This release fixes the privacy-filter remote-code trust boundary, tightens model-name routing for the OpenAI/OpenMed Privacy Filter family, and brings the public HuggingFace-to-MLX converter in line with the already-published OpenMed MLX Privacy Filter artifacts.

Highlights

Hardened Privacy Filter loading so arbitrary Hugging Face repo names containing privacy-filter no longer route through the trusted remote-code path.
Changed PrivacyFilterTorchPipeline so trust_remote_code defaults to False.
Added an explicit allowlist for first-party Privacy Filter repositories that legitimately require Transformers custom-code loading.
Added OPENMED_TRUSTED_REMOTE_CODE_MODELS for operators who need to trust controlled/private fine-tunes.
Fixed raw HuggingFace-to-MLX conversion for the OpenAI Privacy Filter family, including Nemotron and multilingual variants.
Bumped package, README, docs, and website surfaces to 1.5.2.

Security

The Privacy Filter dispatcher previously used broad substring matching for model identifiers. A model name such as attacker/foo-privacy-filter-bar could be treated as part of the privacy-filter family and reach a path where custom repository code may be loaded.

1.5.2 separates routing from trust:

only first-party Privacy Filter identifiers route through the privacy-filter dispatcher;
trust_remote_code=True is only enabled for allowlisted first-party repos or operator-controlled local/env-configured models;
direct PrivacyFilterTorchPipeline(..., trust_remote_code=True) calls now fail fast for untrusted model IDs.

Trusted first-party remote-code models:

openai/privacy-filter
OpenMed/privacy-filter-multilingual
OpenMed/privacy-filter-nemotron

Custom/private fine-tunes can be allowlisted with:

export OPENMED_TRUSTED_REMOTE_CODE_MODELS="my-org/my-privacy-filter-finetune"

What's Changed

Release OpenMed 1.5.1 by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/56
fix: add MLX weight remapping for openai_privacy_filter / nemotron architecture by @kswanjitsu in https://github.com/maziyarpanahi/openmed/pull/58
Harden privacy-filter remote-code allowlist for 1.5.2 by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/59

New Contributors

@kswanjitsu made their first contribution in https://github.com/maziyarpanahi/openmed/pull/58

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.5.1...v1.5.2

Open source →

Additional notes

Security

Hardened the privacy-filter dispatcher to refuse trust_remote_code=True for model identifiers outside an explicit allowlist of first-party OpenAI/OpenMed privacy-filter family models (openai/privacy-filter, OpenMed/privacy-filter-multilingual, OpenMed/privacy-filter-nemotron). Previously, any HuggingFace repository whose name contained the substring privacy-filter would be loaded with custom-code execution enabled, allowing remote code execution by anyone able to control the model_name parameter on /pii/extract or /pii/deidentify. Operators with custom fine-tunes of the privacy-filter family can extend the allowlist via the OPENMED_TRUSTED_REMOTE_CODE_MODELS environment variable (comma-separated repo IDs).
Changed PrivacyFilterTorchPipeline's trust_remote_code default from True to False. The first-party dispatcher (openmed.core.backends.create_privacy_filter_pipeline) opts in explicitly only for allowlisted models.

Changed

README, docs, and website version surfaces now point at 1.5.2.

Fixed

Fixed raw HuggingFace-to-MLX conversion for the OpenAI Privacy Filter family (openai/privacy-filter, OpenMed/privacy-filter-nemotron, and OpenMed/privacy-filter-multilingual) by casting BF16 tensors to float32 before NumPy conversion, remapping OPF/Nemotron checkpoints into the OpenMed MLX runtime layout, fusing Q/K/V projections, preserving classifier bias, and validating converted weight keys/shapes before artifact save.

Tests

Added tests/unit/test_privacy_filter_security.py covering the identifier matcher, allowlist gate, env-var override, local-artifact trust, and dispatcher opt-in.
Added HTTP-level regression tests in tests/unit/service/test_api.py that POST the attacker-controlled model_name payload to /pii/extract and /pii/deidentify and verify the privacy-filter dispatcher is never reached.
Added MLX converter regressions for BF16 NumPy conversion, OPF weight remapping, QKV fusion order, and partial-QKV rejection.

Open source →

1.5.121 May 2026

Patch release for OpenMed 1.5.1.

Changes:

Bumps package metadata to 1.5.1.
Refreshes README, docs, website, and Apple demo version surfaces.
Adds the 1.5.1 changelog entry.

This release tag is intended to trigger the tag-driven build and publish workflows.

Open source →

Additional notes

Changed

README, docs, website, and Apple demo version surfaces now point at 1.5.1.
Prepared the patch release metadata for the tag-driven build and publish workflow.

Open source →

1.5.019 May 2026

OpenMed 1.5.0 expands multilingual PII extraction and MLX availability with Arabic, Japanese, and Turkish support across the Python SDK and preconvert

OpenMed v1.5.0 with Arabic, Japanese, and Turkish support

OpenMed 1.5.0 expands multilingual PII extraction and MLX availability with Arabic, Japanese, and Turkish support across the Python SDK and preconverted Apple MLX artifacts.

Highlights

Added PII extraction/de-identification support for ar, ja, and tr.
Registered all API-visible new-language PII source checkpoints: 2 Arabic, 3 Japanese, and 32 Turkish models.
Increased the SDK PII registry to 247 models.
Added preconverted MLX routing for the 28 supported Arabic/Japanese/Turkish -mlx repositories.
Uploaded and verified all 28 supported MLX repos in the OpenMed Medical MLX Models collection.
Added Turkish TCKN checksum validation and context-aware Arabic/Japanese national ID patterns.
Added anonymizer locale routing for ar_EG, ja_JP, and tr_TR.
Fixed Turkish street-address matching for common name-first forms such as Atatürk Caddesi 12.
Included #53: local model directories now load correctly through model_name or model_id without unintended Hub lookups.

MLX Rollout

The supported Arabic, Japanese, and Turkish PII token-classification models were converted to MLX, uploaded to private OpenMed/<source>-mlx repos, and added to the Medical MLX Models collection.

Converted/uploaded: 28
Added to collection: 28
Collection verification missing: 0
Failed: 0

Deferred models use architectures not covered by the current converter pass: ModernBERT, Qwen3, and Longformer.

SDK Updates

analyze_text(..., model_id=...) now works as an alias for model_name, including existing local model directories.
ModelLoader now preserves filesystem paths before registry/default-org expansion and keeps local-path Hugging Face loading local-only.
extract_pii(..., lang="ar") now defaults to OpenMed/OpenMed-PII-Arabic-SnowflakeMed-Large-568M-v1.
extract_pii(..., lang="ja") now defaults to OpenMed/OpenMed-PII-Japanese-BigMed-Large-560M-v1.
extract_pii(..., lang="tr") now defaults to OpenMed/OpenMed-PII-Turkish-SuperClinical-Small-44M-v1.
OpenMedConfig(backend="mlx") can resolve supported new-language PII checkpoints to uploaded -mlx artifacts directly.

Docs And Release Prep

Bumped package/docs/demo version surfaces to 1.5.0.
Updated the changelog, README, docs, website copy, Swift package snippets, and multilingual smoke example.
Updated the multilingual PII guide from 9 languages / 210 models to 12 languages / 247 models.

Verification

uv run pytest tests/unit/test_pii_i18n.py tests/unit/test_model_registry_multilingual.py tests/unit/mlx/test_mlx_inference.py — 211 passed
uv run pytest tests/unit/test_pii_i18n.py — 150 passed
PR #53 validation: uv run pytest tests/unit/test_core.py tests/unit/test_utils.py -q — 71 passed; uv run pytest tests/unit -q — 1078 passed, 1 skipped
uv run python scripts/release/check_release_version.py --version 1.5.0 — passed
git diff --check — passed

What's Changed

Fix local model path loading by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/53
Release 1.5.0: Arabic, Japanese, and Turkish PII MLX support by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/55

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.4.0...v1.5.0

Open source →

Additional notes

Added

Arabic (ar), Japanese (ja), and Turkish (tr) PII extraction support in the Python SDK, including language defaults, localized regex patterns, fake replacement data, and anonymizer locale routing.
Registry entries for all API-visible Arabic, Japanese, and Turkish PII source checkpoints: 2 Arabic, 3 Japanese, and 32 Turkish models.
Preconverted MLX routing for the 28 supported Arabic, Japanese, and Turkish PII -mlx repositories so OpenMedConfig(backend="mlx") can resolve uploaded artifacts directly.
Turkish TCKN checksum validation plus context-aware Arabic and Japanese national ID patterns.

Changed

README, docs, website, and Apple demo version surfaces now point at 1.5.0.
Faker anonymization now falls back to en_US with a warning if a requested locale is unavailable at runtime.

Fixed

Turkish street-address matching now accepts both descriptor-first forms such as Cadde İnönü 12 and common Turkish name-first forms such as Atatürk Caddesi 12.

Tests

Added language constant/default routing, model registry count, MLX mapping, anonymizer locale, and multilingual PII regression coverage for Arabic, Japanese, and Turkish.

Open source →

1.4.011 May 2026

This release brings the OpenMed Multilingual Privacy Filter into the main OpenMed ecosystem across Python, MLX, OpenMedKit, the iOS Scan Demo, and the

OpenMed v1.4.0 is the multilingual Privacy Filter release.

This release brings the OpenMed Multilingual Privacy Filter into the main OpenMed ecosystem across Python, MLX, OpenMedKit, the iOS Scan Demo, and the web demo experience. The new family officially supports 16 languages and ships in PyTorch, MLX full-precision, and MLX 8-bit forms.

The headline: developers can now use the same extract_pii() / deidentify() API for the OpenAI baseline, OpenAI Nemotron Privacy Filter, and OpenMed Multilingual Privacy Filter, while Apple demos can showcase all three model choices without changing application code.

Highlights

Added the OpenMed Multilingual Privacy Filter model family:
- OpenMed/privacy-filter-multilingual
- OpenMed/privacy-filter-multilingual-mlx
- OpenMed/privacy-filter-multilingual-mlx-8bit
Added Python MLX routing for the multilingual full and 8-bit artifacts.
Added family-aware fallback so multilingual MLX names resolve to the multilingual PyTorch checkpoint on non-MLX hosts.
Added MLX family aliases for multilingual Privacy Filter artifacts that reuse the existing OpenAI Privacy Filter runtime and BIOES decoder.
Updated the OpenMed Scan Demo with the 8-bit multilingual model, a clearer three-model picker, and EN/FR/AR sample buttons.
Added French and Arabic scanned demo documents for screenshot-ready multilingual flows.
Added a multilingual web studio that compares the OpenAI baseline, OpenAI Nemotron Privacy Filter, and OpenMed Multilingual Privacy Filter.
Updated README, anonymization docs, MLX docs, Swift docs, CHANGELOG, and version surfaces for 1.4.0.

Privacy Filter Families

OpenMed now documents and routes three Privacy Filter families:

Variant	PyTorch	MLX full	MLX 8-bit
OpenAI Privacy Filter	`openai/privacy-filter`	`OpenMed/privacy-filter-mlx`	`OpenMed/privacy-filter-mlx-8bit`
OpenAI Nemotron Privacy Filter	`OpenMed/privacy-filter-nemotron`	`OpenMed/privacy-filter-nemotron-mlx`	`OpenMed/privacy-filter-nemotron-mlx-8bit`
OpenMed Multilingual Privacy Filter	`OpenMed/privacy-filter-multilingual`	`OpenMed/privacy-filter-multilingual-mlx`	`OpenMed/privacy-filter-multilingual-mlx-8bit`

All three families use the OpenAI Privacy Filter architecture. The multilingual family uses OpenMed multilingual PII training data and officially supports 16 languages.

Python Usage

The public API stays the same:

from openmed import extract_pii, deidentify

text = "Patient Marie Dubois, nee le 14/03/1982, email marie.dubois@example.fr."

entities = extract_pii(
    text,
    model_name="OpenMed/privacy-filter-multilingual-mlx-8bit",
)

safe = deidentify(
    text,
    model_name="OpenMed/privacy-filter-multilingual-mlx-8bit",
    method="replace",
    consistent=True,
    seed=42,
)

On Apple Silicon with MLX available, the MLX artifact runs through PrivacyFilterMLXPipeline. On other hosts, OpenMed substitutes the matching PyTorch checkpoint and emits a one-time warning:

OpenMed/privacy-filter-mlx* -> openai/privacy-filter
OpenMed/privacy-filter-nemotron-mlx* -> OpenMed/privacy-filter-nemotron
OpenMed/privacy-filter-multilingual-mlx* -> OpenMed/privacy-filter-multilingual

Apple And Demo Updates

The iOS Scan Demo now presents three privacy engines cleanly:

OpenMed PII
OpenAI Nemotron Privacy Filter
OpenMed Multilingual Privacy Filter

The multilingual path uses OpenMed/privacy-filter-multilingual-mlx-8bit so the demo stays aligned with the 8-bit Apple artifact strategy. The sample controls now use compact EN, FR, and AR buttons, and switching language/sample clears previous annotations before the next run starts.

The multilingual web studio now uses a single top-to-bottom scan pass and redacts line by line during that pass, matching the original Privacy Filter Studio demo feel without looping the scan effect.

Upgrade Notes

The package version is now 1.4.0.
Swift demo marketing versions are now 1.4.0.
OpenMed/privacy-filter-multilingual-mlx and OpenMed/privacy-filter-multilingual-mlx-8bit are first-class model names in the MLX routing table.
The multilingual MLX artifacts must include a valid openmed-mlx.json; stale cached HTTP error bodies are no longer treated as manifests by the scan demo downloader.

Validation

This release adds targeted unit coverage for multilingual Privacy Filter routing, MLX family alias dispatch, and family-aware fallback behavior. The OpenMed Scan Demo was also rebuilt after the multilingual 8-bit integration.

What's Changed

Fix strict docs build for anonymization guide by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/48
Add multilingual Privacy Filter v1.4.0 release by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/51

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.3.0...v1.4.0

Open source →

Additional notes

Added

OpenMed Multilingual Privacy Filter family, registered across PyTorch and MLX:
- OpenMed/privacy-filter-multilingual — PyTorch / Transformers (CPU + CUDA).
- OpenMed/privacy-filter-multilingual-mlx — MLX full-precision (Apple Silicon).
- OpenMed/privacy-filter-multilingual-mlx-8bit — MLX 8-bit quantized (Apple Silicon and OpenMedKit demos). These artifacts use the OpenAI Privacy Filter architecture and officially support 16 languages through the OpenMed multilingual PII corpus.
Python MLX routing for multilingual Privacy Filter artifacts:
- _MLX_MODEL_MAP entries for the full and 8-bit multilingual MLX repo IDs.
- privacy-filter-multilingual and multilingual-privacy-filter MLX family aliases, both resolving to the existing OpenAI Privacy Filter model class and BIOES decoder.
- Family-aware Torch fallback so multilingual MLX model names substitute OpenMed/privacy-filter-multilingual on non-MLX hosts instead of the OpenAI baseline.
Multilingual Privacy Filter Studio in examples/privacy_filter_multilingual_studio/, a web demo comparing the OpenAI baseline, OpenAI Nemotron Privacy Filter, and OpenMed Multilingual Privacy Filter with English, French, and Arabic examples.
OpenMed Scan Demo multilingual mode with OpenMed/privacy-filter-multilingual-mlx-8bit, a three-engine picker, EN/FR/AR sample buttons, and new French/Arabic scanned demo documents for screenshot-ready flows.

Changed

Privacy Filter docs and README now describe three Privacy Filter families and label the multilingual model as OpenMed Multilingual Privacy Filter.
OpenMedKit and demo version surfaces now point at 1.4.0.
The scan demo clears previous annotation windows whenever the language/sample changes, avoiding stale entities from earlier model runs.
The multilingual web studio scan animation now performs a single top-to-bottom pass while redacting line by line, matching the stronger visual rhythm of the original Privacy Filter Studio.

Fixed

Improved Swift model-download handling so stale cached 401/404 responses cannot masquerade as openmed-mlx.json manifests after a public model becomes available.
Tightened stale-result invalidation in iOS and web demo flows so slower previous model runs cannot overwrite a newly selected language/sample.

Tests

Added Python unit coverage for multilingual MLX backend selection, family-aware Torch fallback, and MLX Privacy Filter family dispatch aliases.
Rebuilt the OpenMed Scan Demo after the multilingual 8-bit integration.

Open source →

1.3.029 Apr 2026

This release turns PII handling into a more complete cross-platform workflow: Faker-backed obfuscation, deterministic surrogates, a canonical PII labe

#OpenMed v1.3.0 is the privacy and anonymization release.

This release turns PII handling into a more complete cross-platform workflow: Faker-backed obfuscation, deterministic surrogates, a canonical PII label taxonomy, unified Privacy Filter routing across MLX and PyTorch, Nemotron-PII Privacy Filter artifacts, and a new interactive Privacy Filter Studio.

The headline: OpenMed can now detect, mask, remove, hash, date-shift, or realistically replace identifiers with locale-aware surrogates, while using the same extract_pii() / deidentify() API on Apple Silicon, Linux, Windows, and service deployments.

Highlights

Added a Faker-backed anonymization engine for method="replace".
Added deterministic, locale-aware, and format-preserving surrogate generation.
Added a canonical PII label taxonomy that normalizes English, Portuguese, and BIOES-tagged Privacy Filter labels into one stable label set.
Added unified Privacy Filter backend routing: MLX on Apple Silicon, PyTorch everywhere else.
Added PyTorch support for the OpenAI Privacy Filter family through PrivacyFilterTorchPipeline.
Added Nemotron-PII Privacy Filter artifacts for PyTorch and MLX.
Added family-aware fallback so Nemotron MLX model names resolve to the Nemotron PyTorch checkpoint on non-MLX hosts.
Added shared BIOES/Viterbi decoding and span refinement utilities used by both MLX and PyTorch Privacy Filter paths.
Added Portuguese (pt) support to the REST API schemas.
Added Privacy Filter Studio, an interactive FastAPI/static web demo for masking and deterministic randomization.
Added Python and Swift Privacy Filter classifier-head bias support for Nemotron-PII artifacts.

Why This Release Matters

PII de-identification is only useful when the output is both safe and usable.

Simple masking is sometimes enough, but many clinical, operational, and demo workflows need text that still looks realistic: names that look like names, phone numbers that keep their separators, dates that keep their local ordering, and repeated mentions that resolve to the same fake person.

OpenMed v1.3.0 moves beyond static replacement lists. It gives developers a single privacy API that can:

run locally on Apple Silicon through MLX
run on CPU or CUDA through Transformers/PyTorch
preserve downstream-friendly formats
generate locale-appropriate fake identifiers
behave deterministically for reproducible tests and demos
route OpenAI and Nemotron Privacy Filter checkpoints through the same code

That makes OpenMed more practical for clinical prototypes, privacy demos, evaluation harnesses, and local-first healthcare applications.

Faker-Backed Anonymization

method="replace" now uses openmed.core.anonymizer.Anonymizer.

The anonymizer supports:

cached per-locale Faker instances
deterministic seeding with hashlib.blake2b
label-keyed generator dispatch
format-preserving phones, dates, emails, and generic IDs
locale overrides such as locale="pt_BR" or locale="en_GB"
custom label generators through register_label_generator()
custom clinical Faker providers through register_clinical_provider()

Example:

from openmed import deidentify

text = "Patient Pedro Almeida, CPF 123.456.789-09, phone +351 912 345 678."

result = deidentify(
    text,
    method="replace",
    lang="pt",
    locale="pt_BR",
    consistent=True,
    seed=42,
)

print(result.deidentified_text)

Deterministic mode means the same (label, original value) pair maps to the same surrogate within a call. Passing seed= makes the output reproducible across runs.

Clinical And National IDs

OpenMed now includes custom Faker providers for clinical and national ID shapes where Faker's built-ins are missing or insufficient:

Aadhaar with Verhoeff checksum
German Steuer-ID
medical record numbers
US National Provider Identifier (NPI)

It also reuses Faker's locale-specific built-ins where they already validate against OpenMed's checksum logic:

pt_BR.cpf and pt_BR.cnpj
nl_NL.ssn for BSN
fr_FR.ssn for NIR
it_IT.ssn for Codice Fiscale
es_ES.nie

Canonical Label Taxonomy

openmed.core.labels introduces CANONICAL_LABELS and normalize_label().

This gives downstream code one stable label vocabulary even when models emit different naming schemes:

English lowercase labels such as first_name
Portuguese uppercase labels such as FIRSTNAME
Privacy Filter BIOES labels such as B-NAME, I-EMAIL, or S-PHONE
mixed-case or separator variants

The anonymizer, replacement mapping, and Privacy Filter routes now use this normalization layer to reduce model-family-specific branching.

Privacy Filter Family

OpenMed v1.3.0 exposes two Privacy Filter checkpoint families through the same public API:

Variant	PyTorch	MLX full	MLX 8-bit
OpenAI Privacy Filter	`openai/privacy-filter`	`OpenMed/privacy-filter-mlx`	`OpenMed/privacy-filter-mlx-8bit`
Nemotron-PII fine-tune	`OpenMed/privacy-filter-nemotron`	`OpenMed/privacy-filter-nemotron-mlx`	`OpenMed/privacy-filter-nemotron-mlx-8bit`

Both families use the OpenAI Privacy Filter architecture. The Nemotron-PII artifacts are fine-tuned on the Nemotron PII dataset and reuse the existing Privacy Filter pipeline and model architecture.

Use the same API everywhere:

from openmed import extract_pii, deidentify

text = "Patient Sarah Connor, DOB 03/15/1985, MRN 4471882."

entities = extract_pii(
    text,
    model_name="OpenMed/privacy-filter-nemotron-mlx-8bit",
)

safe = deidentify(
    text,
    model_name="OpenMed/privacy-filter-nemotron-mlx-8bit",
    method="replace",
    consistent=True,
    seed=42,
)

On Apple Silicon with MLX available, MLX artifacts run through PrivacyFilterMLXPipeline. On other hosts, MLX-only model names are automatically substituted with the matching PyTorch checkpoint:

OpenMed/privacy-filter-mlx* -> openai/privacy-filter
OpenMed/privacy-filter-nemotron-mlx* -> OpenMed/privacy-filter-nemotron

A one-time UserWarning explains the substitution.

PyTorch Privacy Filter

openmed.torch.PrivacyFilterTorchPipeline loads the Privacy Filter family via Transformers:

auto-selects CUDA when available, otherwise CPU
supports compatible fine-tunes
emits the same entity dictionary shape as the MLX pipeline
uses trust_remote_code=True by default for the OpenAI Privacy Filter family

Install:

pip install -U "openmed[hf]"

Run:

from openmed import extract_pii

result = extract_pii(
    "Alice Smith emailed alice@example.com.",
    model_name="openai/privacy-filter",
)

MLX Privacy Filter Updates

The Python MLX Privacy Filter runtime now shares decoding utilities with the PyTorch path:

TokenLabelInfo
build_label_info
viterbi_decode
labels_to_token_spans
trim_span_whitespace
refine_privacy_filter_span

This keeps BIOES/Viterbi decoding consistent across backends.

The MLX model class also now honors classifier_bias / unembedding_bias in artifact configs. This keeps the original OpenAI Privacy Filter bias-less by default while allowing Nemotron-PII artifacts to load their biased classifier head correctly.

Swift And OpenMedKit

OpenMedKit also gained Privacy Filter classifier-head bias support.

The native MLX artifact loader now decodes classifier_bias / unembedding_bias and builds the Privacy Filter head with a learned bias when Nemotron-PII artifacts require it, while preserving the bias-less baseline path.

The OpenMed Scan Demo privacy-filter option now points at OpenMed/privacy-filter-nemotron-mlx-8bit and labels the engine as OpenAI Nemotron Privacy Filter throughout the picker, download events, and README.

Privacy Filter Studio

This release adds examples/privacy_filter_studio/, an interactive two-pane web demo for PII de-identification.

It includes:

sample clinical and operational notes
mask and deterministic randomize modes
highlighted detected entities
per-entity labels and category colors
model/backend status
latency and entity counters
a first-run download toggle
cache-only model loading unless downloads are explicitly allowed

Run:

pip install -U "openmed[mlx]"        # or "openmed[hf]" off Apple Silicon
uvicorn examples.privacy_filter_studio.app:app --reload --port 8770

Open:

http://127.0.0.1:8770

Override the model:

OPENMED_STUDIO_MODEL=OpenMed/privacy-filter-nemotron-mlx-8bit \
  uvicorn examples.privacy_filter_studio.app:app --port 8770

Documentation And Examples

New and updated docs/examples:

docs/anonymization.md
examples/obfuscation_demo.py
examples/privacy_filter_unified.py
examples/privacy_filter_studio/
examples/privacy_filter_book/app.py

The anonymization guide covers deterministic surrogates, locale resolution, format preservation, custom generators, clinical ID providers, and the Privacy Filter routing model.

Breaking Changes

faker>=22.0 is now a required core dependency.
method="replace" no longer uses the old small static fake-data lists. Downstream tests that asserted exact prior replacement strings should be updated.
Privacy Filter routing through extract_pii() skips regex smart-merging by design, because the model already performs Viterbi-constrained BIOES span construction.

Other de-identification methods are unchanged:

mask
remove
hash
shift_dates

Upgrade Notes

Install or upgrade:

pip install -U openmed

For PyTorch Privacy Filter support:

pip install -U "openmed[hf]"

For Apple Silicon MLX support:

pip install -U "openmed[mlx]"

Recommended checks for application upgrades:

If you assert exact method="replace" outputs, switch to seeded deterministic expectations or assert that originals are removed.
If you use MLX Privacy Filter model names on non-MLX hosts, expect a one-time warning and an automatic PyTorch substitution.
If you package Nemotron-PII MLX artifacts, keep classifier_bias or unembedding_bias in the artifact config when the classifier head has bias.
If you expose downloads in a UI, use explicit user control like Privacy Filter Studio's download toggle.

Validation

Release-prep validation included:

git diff --check
.venv/bin/python -m compileall -q examples/privacy_filter_studio openmed/mlx/models/privacy_filter.py
.venv/bin/python -m pytest tests/unit/mlx/test_privacy_filter_mlx.py tests/unit/test_privacy_filter_routing.py
.venv/bin/python -m pytest tests/unit/core/test_anonymizer.py tests/unit/core/test_labels.py tests/unit/test_pii.py tests/unit/test_privacy_filter_routing.py tests/unit/test_pii_multilingual_regression.py tests/unit/mlx/test_privacy_filter_mlx.py tests/unit/service/test_api.py

Results captured during release prep:

Studio and MLX model compile check: passed
Studio FastAPI smoke test: passed
Privacy Filter routing/MLX subset: 20 passed, 8 skipped
Focused privacy/anonymization suite: 471 passed, 1 skipped, 11 warnings

The warnings are pre-existing span-validation warnings from multilingual PII regression fixtures.

Thank You

OpenMed v1.3.0 is about making privacy work feel less like a demo trick and more like an actual developer surface: local when possible, portable when needed, deterministic when tests demand it, and realistic enough for useful clinical workflows.

Thank you to everyone testing the Privacy Filter artifacts, poking at de-identification edge cases, trying the OpenMedKit paths, and helping OpenMed move toward a more practical open-source healthcare AI stack.

What's Changed

add privacy filter book comparison demo by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/45
Add anonymization, Studio, and Nemotron privacy-filter support by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/46
Refresh website for v1.3 privacy release by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/47

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.2.0...v1.3.0

Open source →

Additional notes

Added

Faker-backed PII anonymization engine (openmed.core.anonymizer):
- Anonymizer class with cached per-locale Faker instances, deterministic seeding (hashlib.blake2b), and label-keyed generator dispatch.
- AnonymizerConfig dataclass for advanced configuration.
- Locale resolution map (LANG_TO_LOCALE) covering all nine OpenMed languages; Telugu falls back to en_IN with a one-time UserWarning.
- Format-preserving helpers for phone numbers (digit-group lengths preserved), dates (separator/ordering preserved), emails (domain preserved), and generic IDs.
- Custom Faker providers for clinical/national IDs where Faker's built-ins are missing or incorrect: AadhaarProvider (Verhoeff checksum), GermanSteuerIdProvider, MedicalRecordNumberProvider, NPIProvider. Faker's built-ins are reused for pt_BR.cpf/cnpj, nl_NL.ssn (BSN), fr_FR.ssn (NIR), it_IT.ssn (Codice Fiscale), and es_ES.nie after empirical verification against OpenMed's existing checksum validators.
- register_clinical_provider() and register_label_generator() for extending coverage.
Canonical PII label taxonomy (openmed.core.labels):
- CANONICAL_LABELS set with 47 canonical labels in UPPER_SNAKE_CASE.
- normalize_label() maps English lowercase, the 52 Portuguese UPPERCASE labels, BIOES-tagged variants (B-NAME, I-DATE), and arbitrary mixed-case forms to a single canonical form.
Unified privacy-filter dispatch (openmed.core.backends):
- select_privacy_filter_backend(), resolve_privacy_filter_model(), and create_privacy_filter_pipeline() route privacy-filter requests to MLX on Apple Silicon and PyTorch elsewhere with a one-time UserWarning when an MLX-only artifact name (OpenMed/privacy-filter-mlx*) is substituted with openai/privacy-filter on non-Mac hosts.
- extract_pii() and deidentify() now route privacy-filter models through this dispatcher, skipping regex smart-merging since the model already does Viterbi-constrained BIOES decoding.
PyTorch privacy-filter wrapper (openmed.torch.PrivacyFilterTorchPipeline):
- Loads openai/privacy-filter (or any compatible HuggingFace fine-tune) via transformers.AutoModelForTokenClassification with auto device selection (CUDA → CPU).
- Output entity-dict shape matches the MLX pipeline so the rest of OpenMed is backend-agnostic.
Shared decoding utilities (openmed.core.decoding):
- TokenLabelInfo, build_label_info, viterbi_decode, labels_to_token_spans, zero_viterbi_biases, VITERBI_BIAS_KEYS extracted from the MLX pipeline so the Torch wrapper reuses the same BIOES Viterbi decoder.
- trim_span_whitespace, refine_privacy_filter_span for span post-processing across both backends.
deidentify() keyword arguments: consistent: bool, seed: Optional[int], locale: Optional[str] for deterministic, locale-overridable obfuscation. Passing seed= alone implies consistent=True.
Portuguese (pt) accepted by REST API schemas in openmed/service/schemas.py (was previously library-only despite full core support).
Documentation: new Anonymization Guide covering the Faker engine, locale table, determinism modes, format preservation, clinical-ID checksum sources, and the privacy-filter family.
Examples:
- examples/obfuscation_demo.py — random vs deterministic surrogates, locale walkthrough, format-preserving phone numbers, pt_BR CPF generation with checksum verification.
- examples/privacy_filter_unified.py — same extract_pii() / deidentify() call works on Apple Silicon (MLX) and Linux (PyTorch); compares the OpenAI baseline against the Nemotron-PII fine-tune side-by-side.
- examples/privacy_filter_studio/ — interactive FastAPI + static web studio for two-pane PII masking/randomization with sample clinical notes, highlighted entities, backend/model status, and an explicit first-run download toggle.
Nemotron-PII fine-tune of the OpenAI Privacy Filter, registered as three new model IDs that route through the existing privacy-filter pipeline (same architecture, different training data):
- OpenMed/privacy-filter-nemotron — PyTorch / Transformers (CPU + CUDA).
- OpenMed/privacy-filter-nemotron-mlx — MLX full-precision (Apple Silicon).
- OpenMed/privacy-filter-nemotron-mlx-8bit — MLX 8-bit quantized (Apple Silicon). These checkpoints are the OpenAI Privacy Filter architecture (gpt-oss-style sparse-MoE transformer with local attention, sink tokens, RoPE+YaRN, tiktoken o200k_base) fine-tuned on the Nemotron PII dataset. They reuse OpenAIPrivacyFilterForTokenClassification and PrivacyFilterMLXPipeline unchanged — no new architecture code needed.
_MLX_MODEL_MAP entries for the two new Nemotron MLX repo IDs in openmed.mlx.inference.
Aliases for the new family in _SUPPORTED_TOKEN_CLASSIFICATION_MODEL_TYPES (privacy-filter-nemotron, nemotron-privacy-filter) — both resolve to the existing openai-privacy-filter family so a Nemotron-fine-tune MLX artifact can ship with either family identifier in its manifest and still dispatch correctly.
Family-aware Torch fallback in openmed.core.backends:
- New _TORCH_FALLBACK_BY_FAMILY table and _torch_fallback_for() helper.
- An MLX-only Nemotron request on a non-Apple-Silicon host now substitutes OpenMed/privacy-filter-nemotron instead of the unrelated default openai/privacy-filter, so the user gets the training distribution they asked for. A one-time UserWarning names the substitute.
- Adding a future fine-tune that should fall back to its own PyTorch repo is a one-line addition to _TORCH_FALLBACK_BY_FAMILY.
Nemotron MLX classifier-head bias support: OpenAIPrivacyFilterForTokenClassification now honors classifier_bias / unembedding_bias in artifact configs, while keeping the original OpenAI checkpoint bias-less by default.
Swift OpenMedKit privacy-filter classifier-head bias support: the native MLX artifact loader now decodes classifier_bias / unembedding_bias and builds the Privacy Filter head with a learned bias when Nemotron-PII artifacts require it.

Changed

method="replace" upgraded in place to use the new Faker-backed Anonymizer. Surrogates are now locale-aware (e.g. German names for lang="de", Portuguese phones for lang="pt"), format-preserving, and optionally deterministic. The previous tiny static LANGUAGE_FAKE_DATA lists are kept as a deprecated fallback used only when a Faker locale is unavailable.
Privacy filter book demo (examples/privacy_filter_book/app.py) migrated to PrivacyFilterTorchPipeline for the CPU side, replacing the inline AutoTokenizer/AutoModelForTokenClassification/pipeline triple.
MLX inference module trimmed: BIOES Viterbi (≈280 lines) and span-refinement helpers moved to openmed.core.decoding. Behavior unchanged.
Privacy Filter Studio keeps model loading cache-only unless downloads are explicitly allowed, then restores the caller's Hugging Face offline environment after loading.
OpenMed Scan Demo privacy-filter option now points at OpenMed/privacy-filter-nemotron-mlx-8bit and labels the engine as OpenAI Nemotron Privacy Filter throughout the picker, download events, and README.

Breaking Changes

faker>=22.0 is now a required core dependency. Slim installs that skip the ML extras will still pull Faker (~3 MB).
method="replace" outputs no longer come from the prior hardcoded list (["Jane Smith", "John Doe", "Alex Johnson", "Sam Taylor"], etc.). Any test or downstream code asserting on those exact strings must either pass consistent=True, seed=<value> and update expected output, or assert non-equality with the original. All other methods (mask, remove, hash, shift_dates) are unchanged.
Privacy-filter routing through extract_pii() skips regex smart-merging by design. Users who previously chained the low-level MLX pipeline with merge_entities_with_semantic_units() manually may see different entity counts; the new path produces cleaner spans because the model's Viterbi decoder already enforces BIOES validity.

Tests

New tests across tests/unit/core/test_labels.py (102), tests/unit/core/test_anonymizer.py (171, includes per-locale checksum validation across 100s of generated IDs), tests/unit/test_privacy_filter_routing.py (22 — backend selection, family-aware Torch fallback, dispatch, integration), Nemotron parametrisation of the existing privacy-filter MLX dispatch test (tests/unit/mlx/test_privacy_filter_mlx.py::test_dispatches_privacy_filter_pipeline), and Portuguese obfuscation regressions in tests/unit/test_pii_multilingual_regression.py (3).
Swift OpenMedKit coverage for classifier_bias / unembedding_bias config decoding, Nemotron-biased Privacy Filter forward shape, and the baseline bias-less head.
Focused privacy/anonymization suite: 458 passed, 6 skipped, 11 pre-existing span-validation warnings.

1.2.024 Apr 2026Breaking

No intentional Python API breaking changes from v1.0.0.

OpenMed v1.2.0 is the on-device healthcare AI release.

This release takes the Apple and MLX foundation from v1.0.0 and turns it into a fuller local inference platform: Python MLX, Swift OpenMedKit, public Privacy Filter artifacts, experimental GLiNER-family extraction, and a redesigned iPhone scan demo that feels like a product instead of a debug console.

The headline: OpenMed can now run more of the clinical document workflow locally, from PII de-identification to clinical entity and relation extraction, across Python, macOS, and iOS.

Highlights

Expanded Python MLX runtime support for OpenMed token-classification artifacts, OpenAI Privacy Filter artifacts, GLiNER span NER, GLiClass zero-shot classification, and GLiNER-Relex relation extraction.
Added native Swift OpenMedKit MLX runtimes for DeBERTa-v2/v3-backed GLiNER-family artifacts and OpenAI Privacy Filter.
Added public Swift APIs for zero-shot NER, zero-shot classification, and relation extraction.
Added local support for OpenMed/privacy-filter-mlx and OpenMed/privacy-filter-mlx-8bit, including the smaller 8-bit artifact for Apple apps.
Rebuilt the scan demo into a guided iPhone workflow: document input, OCR review, PII de-identification, clinical extraction, summary, comparison, and export.
Improved model download and cache behavior so public artifacts are prepared once and reused offline.
Removed Hugging Face token UI from demo flows now that release artifacts are public.

Why This Release Matters

Healthcare AI demos often look impressive until the document leaves the device.

OpenMed v1.2.0 moves in the opposite direction. The goal is a practical local-first workflow where a clinical note can be scanned, reviewed, de-identified, analyzed, and summarized without requiring inference through an external service.

That matters for:

clinical app prototypes that need to feel real on iPhone and macOS
teams evaluating local PII and de-identification workflows
developers building Swift apps around OpenMed models
researchers comparing OpenMed PII, OpenAI Privacy Filter, and GLiNER-style extraction
anyone who wants healthcare NLP to be portable, inspectable, and less dependent on vendor-hosted inference

Python MLX

Python MLX now understands OpenMed custom task artifacts through openmed-mlx.json manifests.

New or expanded runtime paths include:

token-classification
openai-privacy-filter
zero-shot-ner
zero-shot-sequence-classification
zero-shot-relation-extraction

The Privacy Filter runtime includes:

tiktoken-style tokenization
byte-offset reconstruction
BIOES/Viterbi decoding
model-led span repair
support for weights.safetensors with weights.npz fallback

Install:

pip install -U "openmed[mlx]"

Run the public 8-bit Privacy Filter artifact:

from huggingface_hub import snapshot_download
from openmed.mlx.inference import create_mlx_pipeline

model_path = snapshot_download("OpenMed/privacy-filter-mlx-8bit")
pipe = create_mlx_pipeline(model_path)

entities = pipe("Alice Smith emailed alice@example.com and called 415-555-0101.")
print(entities)

Run GLiNER-Relex relation extraction:

from huggingface_hub import snapshot_download
from openmed.mlx.inference import GLiNERRelexMLXPipeline

model_path = snapshot_download("OpenMed/gliner-relex-base-v1.0-mlx")
extractor = GLiNERRelexMLXPipeline(model_path)

result = extractor.inference(
    "Aspirin was prescribed for headache after the patient reported migraine symptoms.",
    labels=["medication", "condition", "symptom"],
    relations=["treats", "associated with"],
    threshold=0.5,
    relation_threshold=0.9,
)

print(result["entities"])
print(result["relations"])

Swift And OpenMedKit

OpenMedKit now goes well beyond the first token-classification milestone.

New Swift runtime support includes:

native DeBERTa-v2/v3 MLX modules for GLiNER-family artifacts
native OpenAI Privacy Filter MLX model support
native Privacy Filter tokenizer and post-processing
OpenMed MLX artifact task/family validation
cached public model downloads through OpenMedModelStore

New public APIs:

OpenMedZeroShotNER
OpenMedZeroShotClassifier
OpenMedRelationExtractor

Swift Package Manager:

dependencies: [
    .package(url: "https://github.com/maziyarpanahi/openmed.git", from: "1.2.0"),
]

Run Privacy Filter on device:

import OpenMedKit

let modelURL = try await OpenMedModelStore.downloadMLXModel(
    repoID: "OpenMed/privacy-filter-mlx-8bit"
)

let openmed = try OpenMed(backend: .mlx(modelDirectoryURL: modelURL))

let entities = try openmed.extractPII(
    "Alice Smith emailed alice@example.com and called 415-555-0101."
)

Run GLiNER-Relex relation extraction:

import OpenMedKit

let modelURL = try await OpenMedModelStore.downloadMLXModel(
    repoID: "OpenMed/gliner-relex-base-v1.0-mlx"
)

let extractor = try OpenMedRelationExtractor(modelDirectoryURL: modelURL)

let result = try extractor.extract(
    "Aspirin was prescribed for headache after migraine symptoms.",
    entityLabels: ["medication", "condition", "symptom"],
    relationLabels: ["treats", "associated with"],
    threshold: 0.5,
    relationThreshold: 0.9
)

Demo Apps

The demo apps are now much closer to the release experience we want users to see.

swift/OpenMedDemo now includes the public 8-bit Privacy Filter artifact as a selectable MLX model for macOS and iOS testing.

swift/OpenMedScanDemo has been redesigned around a guided clinical workflow:

Add a document by scanning, loading the sample, or pasting text.
Review OCR output before analysis.
Run de-identification with a selectable PII engine.
Run clinical extraction with GLiNER-Relex presets.
Review safe text, identifiers removed, clinical entities, and export-ready results.

The scan demo also includes:

a generated clinical sample document image
model preparation and download progress UI
offline cache reuse after first download
PII engine comparison
App Store privacy assets for local document scanning
no visible token setup flow in the public release path

Upgrade Notes

No intentional Python API breaking changes from v1.0.0.
Update Swift package requirements to from: "1.2.0" after the release tag is published.
MLX inference should be validated on Apple Silicon macOS or a physical iPhone/iPad.
iOS Simulator remains outside the MLX acceptance path.
GLiNER-family MLX support is experimental while parity and quality testing continue.
Conversion/export internals remain active platform work. Public users should consume the published OpenMed MLX artifacts.

Validation

Release-prep validation completed on April 24, 2026:

python -m pytest tests/unit/mlx/test_mlx_inference.py tests/unit/mlx/test_privacy_filter_mlx.py tests/unit/test_pii.py tests/unit/test_pii_entity_merger.py
cd swift/OpenMedKit && swift test
xcodebuild -project swift/OpenMedDemo/OpenMedDemo.xcodeproj -scheme OpenMedDemo -destination 'generic/platform=iOS' build
xcodebuild -project swift/OpenMedScanDemo/OpenMedScanDemo.xcodeproj -scheme OpenMedScanDemo -destination 'generic/platform=iOS' build

Results:

Python targeted tests: 143 passed, 1 skipped
Swift OpenMedKit tests: 47 passed, 12 skipped
OpenMedDemo generic iOS build: passed
OpenMedScanDemo generic iOS build: passed

The Swift skips are the existing SwiftPM CLI guard for tests that require real MLX runtime resources. Physical-device MLX smoke testing remains the acceptance path for those gated artifact tests.

Thank You

OpenMed v1.2.0 is a release about making healthcare NLP feel more local, more practical, and more usable by real app developers.

Thank you to everyone testing the models, trying the Apple demos, filing sharp feedback, and pushing OpenMed toward a better open-source healthcare AI stack.

What's Changed

Redesign marketing website with editorial design system by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/41
Release 1.1.0 Apple MLX, Privacy Filter, and Scan Demo by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/43
Add 1.2.0 release version preflight by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/44

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.1.0...v1.2.0

Open source →

Additional notes

Added

Expanded Python MLX runtime support for OpenMed MLX artifacts beyond classic token classification, including GLiNER span NER, GLiClass zero-shot classification, GLiNER-Relex relation extraction, and OpenAI Privacy Filter artifacts.
Native OpenAI Privacy Filter MLX pipeline with tiktoken-compatible tokenization, byte-offset reconstruction, BIOES/Viterbi decoding, model-led span repair, and support for the public OpenMed/privacy-filter-mlx and OpenMed/privacy-filter-mlx-8bit artifacts.
Native Swift OpenMedKit GLiNER-family APIs:
- OpenMedZeroShotNER
- OpenMedZeroShotClassifier
- OpenMedRelationExtractor
Native Swift MLX DeBERTa-v2/v3 and Privacy Filter runtimes for local inference on Apple Silicon macOS and physical iPhone/iPad devices.
Self-contained OpenMed MLX artifact handling for task/family manifests, tokenizer assets, weights.safetensors, and weights.npz fallback paths.
OpenMed Scan Demo: a guided iPhone workflow for document capture/sample loading, OCR review, PII de-identification, clinical extraction, summary review, model preparation, and PII engine comparison.
OpenMedDemo Privacy Filter option so macOS/iOS users can test the public OpenAI Privacy Filter MLX artifact alongside OpenMed PII models.
App privacy readiness assets for the scan demo, including a privacy manifest and camera usage copy for local document scanning.

Changed

Improved Apple model download/caching behavior so MLX artifacts are prepared once and reused offline from cache.
Removed Hugging Face token UI and token persistence from demo flows now that release artifacts are public.
Updated PII post-processing so Privacy Filter regex logic repairs model-predicted spans without inventing unsupported semantic labels.
Refreshed OpenMedKit documentation and examples for native MLX artifacts, Swift package usage, and on-device Apple workflows.

Fixed

Reduced iOS memory pressure in the Privacy Filter MLX loader by tightening the Swift model loading path.
Fixed local MLX artifact loading and model-store readiness checks for public Hub artifacts.
Tightened PII entity merging and privacy-filtering tests around model/pattern span interactions.

Tests

Added Python unit coverage for MLX custom-task dispatch, Privacy Filter inference/decoding, artifact loading, and PII privacy-filter post-processing.
Added Swift unit coverage for MLX artifact validation, DeBERTa/GLiNER-family runtime setup, Privacy Filter decoding, sample OCR assets, and post-processing behavior.

1.1.021 Apr 2026

OpenMed v1.1.0 adds first-class Portuguese PII detection and de-identification support with lang="pt".

Portuguese PII & De-identification

OpenMed v1.1.0 adds first-class Portuguese PII detection and de-identification support with lang="pt".

Highlights

Added Portuguese as the 9th supported PII language.
Registered 31 API-visible public Portuguese PII models from the OpenMed Hugging Face collection.
Default Portuguese model: OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1
Expanded the multilingual PII catalog from 179 to 210 registered PII models.
Added Portuguese-aware regex and semantic matching for:
- Dates: 15/03/1985, 15-03-1985, 15 de março de 1985
- Phones: Portugal and Brazil formats
- CPF and CNPJ with checksum validation
- Street addresses
- Portuguese and Brazilian postcodes
Added Portuguese fake replacement data for de-identification.
Updated docs, examples, website copy, and multilingual notebooks.

Quick Start

uv pip install "openmed[hf]"

from openmed import extract_pii

text = (
    "Dr. Pedro Almeida, CPF: 123.456.789-09, "
    "email: pedro@hospital.pt, tel: +351 912 345 678"
)

result = extract_pii(text, lang="pt")

print([(entity.label, entity.text) for entity in result.entities])

Expected detection includes name, CPF, email, and phone entities.

Validation

Tested with targeted unit, registry, multilingual regression, label-normalization, and smoke-example coverage.

369 targeted Portuguese/multilingual tests passed
47 PII accuracy tests passed
Live Portuguese extraction verified against the default public model

What's Changed

Add support for Portuguese PII models by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/42

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.0.0...v1.1.0

Open source →

Additional notes

Added

Portuguese PII and de-identification support via lang="pt"
- Registered 31 API-visible Portuguese PII checkpoints from the OpenMed Hugging Face collection
- Default Portuguese model: OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1
- Added Portuguese regex/semantic patterns for dates, phones, CPF, CNPJ, street addresses, and postcodes
- Added CPF and CNPJ checksum validators, Portuguese fake replacement data, and localized date shifting
Portuguese docs and examples
- Updated multilingual PII documentation from 8 to 9 languages and from 179 to 210 PII models
- Added a Portuguese model-card/README one-liner and smoke-example coverage

Changed

Expanded PII label normalization and replacement mapping for CPF/CNPJ and Portuguese model labels.

Open source →

1.0.014 Apr 2026

This version brings together the core Python runtime, Apple Silicon MLX support, a public Swift package, and a much clearer Apple-platform story.

Highlights

Stable openmed package for Python workflows
Apple Silicon MLX backend for supported token-classification models
New OpenMedKit Swift package for macOS and iOS
Native Xcode demo app
safetensors-first MLX packaging with weights.npz fallback
Improved docs across Python, Swift, MLX, and Apple runtime usage

Apple / MLX Support

OpenMed 1.0.0 includes MLX runtime support across:

BERT
DistilBERT
RoBERTa
XLM-RoBERTa
ELECTRA
DeBERTa-v2 / DeBERTa-v3-style checkpoints

Swift

OpenMedKit is now part of the public 1.0.0 developer story for Apple apps.

Use it from Xcode as a Swift package to integrate OpenMed into macOS and iOS projects.

Demo App

The included Xcode demo now defaults to:

OpenMed/OpenMed-PII-LiteClinical-Small-66M-v1

Notes

Universal Apple packaging across the broader OpenMed model collection is still being expanded.
Some model families are not yet part of the MLX rollout.
The public runtime and developer experience are the main 1.0.0 milestone.

What's Changed

v0.7.0: MLX hardware-accelerated inference for macOS and iOS by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/34
v1.0.0: MLX hardware-accelerated inference for macOS and iOS by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/35
Release 1.0.0 Apple runtime, Swift package, and docs by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/39

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.6.4...v1.0.0

Open source →

Additional notes

Added

Apple MLX inference backend for hardware-accelerated NER on Apple Silicon
- openmed.mlx.models.bert_tc: Pure MLX BERT implementation with token-classification head
- openmed.mlx.inference: MLX NER pipeline producing HuggingFace-compatible output format
- openmed.mlx.convert: CLI tool to convert HuggingFace token-classification models to MLX format with optional 4/8-bit quantization
- Supports BIO tag decoding with simple, first, average, and max aggregation strategies
- Auto-detection: prefers MLX on Apple Silicon when available, falls back to HuggingFace/PyTorch
CoreML export for iOS and macOS deployment
- openmed.coreml.convert: CLI tool to convert HuggingFace models to CoreML .mlpackage format
- Supports flexible sequence lengths via ct.RangeDim, float16/float32 precision
- Embeds id2label mapping in model metadata for self-contained deployment
Swift package: OpenMedKit (swift/OpenMedKit/)
- SPM package for iOS 16+ / macOS 13+ with CoreML-based NER inference
- NERPipeline: CoreML inference with softmax → BIO decoding → entity extraction
- PostProcessing: BIO tag grouping with first/average/max aggregation strategies
- EntityPrediction: Swift equivalent of Python's EntityPrediction dataclass
- Uses swift-transformers for HuggingFace-compatible tokenization
- Includes unit tests for BIO decoding and aggregation strategies
Backend abstraction layer (openmed.core.backends)
- InferenceBackend protocol with is_available() and create_pipeline() interface
- HuggingFaceBackend and MLXBackend implementations
- get_backend() auto-detection with explicit override via config.backend
New optional dependency groups: pip install openmed[mlx] and pip install openmed[coreml]
Pilot model: OpenMed-PII-SuperClinical-Small-44M-v1 as conversion and testing target
37 new tests for backends, MLX conversion key remapping, MLX pipeline output format, CoreML module structure

Changed

Added backend field to OpenMedConfig (None/auto, "hf", "mlx")

Documentation

Updated README, CHANGELOG, and website for the v1.0.0 release

Open source →

0.6.427 Mar 2026

v0.6.4 fixes the issues surfaced by the v0.6.3 quality gates: tokenizer span extension now handles Unicode combining marks and diacritics, pattern val

OpenMed v0.6.4 — Multilingual PII Accuracy & Span-Alignment Consistency

Release date: 2026-03-24

v0.6.4 fixes the issues surfaced by the v0.6.3 quality gates: tokenizer span extension now handles Unicode combining marks and diacritics, pattern validation failures are properly penalized in confidence scoring, and overly-permissive regex patterns for postal codes, phone numbers, and national IDs have been tightened across French, German, Hindi, and Telugu.

What's New

Aadhaar National ID Support (Hindi & Telugu)

New Verhoeff checksum validator and pattern for Indian Aadhaar numbers:

12-digit format with optional 4-4-4 spacing
Rejects invalid prefixes (0xxx, 1xxx) and checksum failures
Context-aware scoring with Hindi (आधार) and Telugu (ఆధార్) keywords

Unicode-Aware Span Extension

_fix_entity_spans now correctly handles non-ASCII text:

Replaced .isalnum() with unicodedata.category check covering letters (L), combining marks (M), and numbers (N)
Capped forward extension at 10 characters to prevent runaway spans
Removed redundant .strip() that created text-mismatch false positives in the quality gate

Relaxed Quality Gate Text-Mismatch

Whitespace-only differences between text[start:end] and entity.text (common after span normalization) are now downgraded from WARNING to INFO level. Genuine text mismatches remain WARNING + SpanValidationWarning.

Validation-Aware Confidence Scoring

When a pattern's validator fails (e.g., SSN checksum, NIR key), merged confidence now uses a 90/10 model/pattern weight (instead of the normal 60/40). This prevents high-confidence scores on structurally-invalid entities.

Pattern Tightening

Pattern	Before	After	Impact
French postal code	`\d{5}`	`01-95 + 971-976` prefixes	Rejects medical codes, invalid dept prefixes
German Steuer-ID	`\d{11}`	`[1-9]\d{10}`	Rejects leading-zero sequences
German postal code	`\d{5}`	`01xxx-99xxx`	Rejects `00xxx` range
German phone	`\d{2,4}[\s/-]?\d{3,8}`	`\d{2,4}[\s/-]?\d{4,8}`	Rejects short (< 4 digit) suffixes

Confidence Calibration

Pattern	Old base_score	New base_score	Rationale
French NIR	0.40	0.55	High structural specificity + validator
German Steuer-ID	0.20	0.35	Tightened pattern + validator
French postal code	0.30	0.25	Still ambiguous even with prefix filter
German postal code	0.30	0.25	Same reasoning

Label Normalization Expansion

New normalize_label() mappings:

bsn, dni, nie, aadhaar → national_id
medical_record_number, mrn → medical_record
account_number → account
credit_debit_card, credit_card, debit_card → payment_card

Test Summary

Suite	Tests	Status
Span-boundary guards	21	All pass
PII accuracy	28	All pass
Multilingual regression	33	All pass
Label-map consistency	46	All pass
Full suite	660	All pass

Files Added

tests/unit/test_pii_accuracy.py — Confidence penalty, pattern tightening, and calibration tests

Files Changed

openmed/processing/outputs.py — Unicode-aware _fix_entity_spans with capped extension
openmed/core/quality_gates.py — Relaxed text-mismatch with whitespace fallback
openmed/core/pii_entity_merger.py — Validation flag in merging, expanded normalize_label
openmed/core/pii_i18n.py — Aadhaar validator + patterns, tightened postal/phone/ID patterns, score calibration
openmed/__about__.py — Version 0.6.3 → 0.6.4
docs/website/index.html — softwareVersion → 0.6.4
CHANGELOG.md — Added v0.6.4 section
README.md — Updated version references
tests/unit/test_quality_gates.py — Combining-mark and whitespace-mismatch tests
tests/unit/test_pii_multilingual_regression.py — Aadhaar tests for Hindi/Telugu
tests/unit/ner/test_label_map_consistency.py — Expanded normalize_label coverage
tests/unit/test_pii_entity_merger.py — Updated for 6-element semantic unit tuples

What's Changed

v0.6.4: multilingual PII accuracy and span-alignment consistency by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/32

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.6.3...v0.6.4

Open source →

Additional notes

Added

Aadhaar national ID support for Hindi and Telugu PII detection
- Added Verhoeff checksum validator (validate_aadhaar) for 12-digit Aadhaar numbers
- Added Aadhaar patterns with context-aware scoring to Hindi and Telugu pattern libraries
PII accuracy test suite (tests/unit/test_pii_accuracy.py)
- Validation-failure confidence penalty tests
- Pattern tightening regression tests (postal codes, phone numbers, Steuer-ID)
- Confidence calibration verification
- New normalize_label coverage tests

Changed

_fix_entity_spans now Unicode-aware — replaced .isalnum() with unicodedata.category check covering letters, combining marks, and numbers; capped forward extension at 10 characters; removed redundant .strip() that caused text-mismatch false positives
Quality gate text-mismatch relaxed — whitespace-only differences (common after span normalization) are now downgraded to INFO level instead of WARNING
Failed pattern validation now penalized in merged confidence — unvalidated patterns contribute only 10% weight (down from 40%) in the model/pattern confidence blend
normalize_label expanded with bsn, dni, nie, aadhaar → national_id; mrn → medical_record; account_number → account; credit_debit_card → payment_card

Fixed

French postal code pattern tightened from bare \d{5} to range-constrained 01-95 + DOM-TOM 971-976 prefixes — reduces false positives from medical codes
German Steuer-ID pattern tightened to reject leading-zero numbers ([1-9]\d{10}); base_score raised to 0.35
German postal code pattern tightened to exclude 00xxx range
German phone pattern now requires at least 4 digits after area code, reducing short-number false positives
French NIR base_score raised from 0.4 to 0.55 to reflect high structural specificity with validator

Documentation

Updated README, CHANGELOG, and website for the v0.6.4 release

Open source →

0.6.319 Mar 2026

v0.6.3 hardens the PII/NER extraction pipeline with deterministic guardrails: span-boundary validation, multilingual regression coverage, and label-ma

OpenMed v0.6.3 — PII/NER Quality Gate Pack

Release date: 2026-03-19

v0.6.3 hardens the PII/NER extraction pipeline with deterministic guardrails: span-boundary validation, multilingual regression coverage, and label-map consistency checks. These quality gates catch tokenizer bugs, model drift, and label inconsistencies before they reach users.

What's New

Span-Boundary Quality Gates

New runtime validation module (openmed.core.quality_gates) that runs automatically after tokenizer repair and smart merging:

validate_entity_spans(entities, text) — checks every entity for:
- start < end (no inverted or zero-length spans)
- Bounds within text (start >= 0, end <= len(text))
- text[start:end] matches stored entity text (catches stale spans after merging)
detect_overlapping_entities(entities) — returns overlapping span pairs for informational use
Warn-only by design — emits SpanValidationWarning and tags entity.metadata["span_valid"], but never silently drops entities

Integrated into both OutputFormatter.format_predictions() (after _fix_entity_spans) and extract_pii() (after smart merging).

Multilingual PII Regression Test Suite

31 golden-input regression tests covering all 8 supported languages:

Language	Tests	Entity types validated
English (en)	5	NAME, DATE, PHONE, SSN, merging
French (fr)	4	NAME, DATE, PHONE, NIR
German (de)	4	NAME, DATE, PHONE, merging
Italian (it)	3	NAME, DATE, PHONE
Spanish (es)	4	NAME, DATE, PHONE, confidence
Dutch (nl)	3	NAME, DATE, PHONE
Hindi (hi)	4	NAME, DATE, PHONE, merging
Telugu (te)	4	NAME, DATE, PHONE, confidence

All tests use mocked model output for fast, deterministic execution.

Label-Map Consistency Tests

32 tests validating configuration invariants:

Every domain in defaults.json has at least 1 label, no case-insensitive duplicates
generic fallback domain always exists
normalize_label() is idempotent for all known label variants
Specificity hierarchy (is_more_specific()) agrees with documented hierarchies
All entity_types in OPENMED_MODELS PII entries are recognized by normalize_label()
At least one PII model per supported language in the registry

Span-Boundary Guard Tests

19 unit tests covering valid entities, inverted/zero-length spans, negative start, out-of-bounds end, text mismatch, overlap detection (adjacent, nested, multiple), and integration with _fix_entity_spans output.

Changed

Website model count updated from 640+ to 750+

Test Summary

Suite	Tests	Status
Span-boundary guards	19	All pass
Multilingual PII regression	31	All pass
Label-map consistency	32	All pass
Full suite	600	All pass

Files Added

openmed/core/quality_gates.py — Span validation + overlap detection
tests/unit/test_quality_gates.py — Guard unit tests
tests/unit/test_pii_multilingual_regression.py — Multilingual regression tests
tests/unit/ner/test_label_map_consistency.py — Label-map invariant tests

Files Changed

openmed/processing/outputs.py — Integrated span guard after _fix_entity_spans
openmed/core/pii.py — Integrated span guard after smart merging
openmed/__about__.py — Version 0.6.2 → 0.6.3
docs/website/index.html — softwareVersion → 0.6.3
CHANGELOG.md — Added v0.6.3 section
README.md — Updated version references

What's Changed

v0.6.3: PII/NER quality gate pack by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/31

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.6.2...v0.6.3

Open source →

Additional notes

Added

Span-boundary quality gates (openmed.core.quality_gates)
- validate_entity_spans() checks start < end, in-bounds, text-match, and zero-length invariants for every entity after tokenizer repair and smart merging
- detect_overlapping_entities() returns pairs of overlapping character spans for informational use
- SpanValidationWarning emitted on violations — warn-only, never silently drops entities
- Integrated into OutputFormatter.format_predictions() (after _fix_entity_spans) and extract_pii() (after smart merging)
Multilingual PII regression test suite (tests/unit/test_pii_multilingual_regression.py)
- Golden-input regression tests for all 8 supported languages (en, fr, de, it, es, nl, hi, te)
- Validates entity type detection, span text matching, confidence thresholds, and smart merging boundaries
- 31 deterministic test cases using mocked model output
Span-boundary guard tests (tests/unit/test_quality_gates.py)
- 19 tests covering valid entities, inverted/zero-length spans, out-of-bounds, text mismatch, overlap detection, and integration with _fix_entity_spans
Label-map consistency tests (tests/unit/ner/test_label_map_consistency.py)
- Validates defaults.json domain invariants (at least 1 label per domain, no case-insensitive duplicates, generic domain exists)
- normalize_label() idempotency checks across all known label variants
- Specificity hierarchy validation against is_more_specific()
- All PII entity_types in OPENMED_MODELS recognized and idempotent under normalize_label()
- At least one PII model per supported language in the registry

Changed

Updated website model count from 640+ to 750+

Documentation

Updated README, website copy, and CHANGELOG for the v0.6.3 release

Open source →

0.6.210 Mar 2026

v0.6.2 expands OpenMed's multilingual PII support and hardens the new REST service for production use.

OpenMed v0.6.2

v0.6.2 expands OpenMed's multilingual PII support and hardens the new REST service for production use.

Highlights

Added public Dutch, Hindi, and Telugu PII models:
- OpenMed/OpenMed-PII-Dutch-SuperClinical-Large-434M-v1
- OpenMed/OpenMed-PII-Hindi-SuperClinical-Large-434M-v1
- OpenMed/OpenMed-PII-Telugu-SuperClinical-Large-434M-v1
Expanded multilingual PII coverage to 179 models across 8 languages
Added Dutch BSN validation plus Dutch/Hindi/Telugu locale patterns, fake data, and localized date handling
Added explicit Python examples and a new multilingual smoke example for the three newly public languages
Hardened the REST service with shared runtime/model reuse, optional preload, stricter request validation, timeout handling, and a unified error envelope

Multilingual PII

extract_pii() and deidentify() now support:

lang="nl" for Dutch
lang="hi" for Hindi
lang="te" for Telugu

French, German, Italian, and Spanish continue to expose the full 35-model multilingual family. Dutch, Hindi, and Telugu currently ship one flagship public checkpoint each.

REST Service

The FastAPI service introduced in v0.6.1 now includes:

shared ServiceRuntime and ModelLoader reuse across requests
optional startup warm-up via OPENMED_SERVICE_PRELOAD_MODELS
threadpool-backed blocking inference with profile timeout enforcement
consistent JSON error envelopes for validation, bad request, timeout, and internal error cases
stricter schema validation for model names, confidence thresholds, extra fields, and shift_dates

Documentation and Examples

README updated for v0.6.2
website copy updated to reflect 8 languages / 179 PII models
REST service guide updated with preload and error-envelope behavior
multilingual notebook refreshed for Dutch, Hindi, and Telugu
added examples/pii_multilingual_new_languages.py

Validation

Focused suite: 306 passed in 0.44s
Live smoke reviewed on the three public checkpoints
- Dutch: name, DOB, BSN, phone, address, postcode, city
- Hindi: date, phone, street address, PIN
- Telugu: date, phone, PIN, with regex/semantic merging covering the sample address span

What's Changed

v0.6.2: harden REST service and add Dutch/Hindi/Telugu PII support by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/29

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.6.1...v0.6.2

Open source →

Additional notes

Added

Dutch, Hindi, and Telugu PII support
- extract_pii() and deidentify() now accept lang="nl", lang="hi", and lang="te"
- Added sparse public registry entries for:
  - OpenMed/OpenMed-PII-Dutch-SuperClinical-Large-434M-v1
  - OpenMed/OpenMed-PII-Hindi-SuperClinical-Large-434M-v1
  - OpenMed/OpenMed-PII-Telugu-SuperClinical-Large-434M-v1
- Added locale-aware patterns for Dutch BSN, Dutch postcodes, India PIN codes, localized month names, and day-first date shifting
- Added Dutch BSN checksum validation and locale-specific fake replacement data for nl, hi, and te
- Added examples/pii_multilingual_new_languages.py for registry, regex, and live-model smoke coverage
REST service runtime hardening
- Added openmed.service.runtime.ServiceRuntime for shared per-process config and model-loader reuse
- Added OPENMED_SERVICE_PRELOAD_MODELS to warm selected models at startup
- Added structured validation/bad-request/timeout/internal-error JSON envelopes for non-2xx responses
- Added request timeout enforcement around blocking inference work
Testing coverage
- Added regression tests for Dutch, Hindi, and Telugu routing, patterns, fake data, date handling, and entity merging
- Added REST service tests for validation errors, timeout behavior, shared-loader reuse, preload parsing, and the new lang values

Changed

Expanded the multilingual PII catalog from 176 to 179 models across 8 languages
get_pii_models_by_language() now returns sparse public releases for nl, hi, and te while keeping English filtering correct
ModelLoader.create_pipeline() now caches created pipelines for repeated requests with identical parameters
REST schemas now validate model names, confidence thresholds, extra fields, and the legacy shift_dates alias more strictly
Updated multilingual examples, notebook guidance, website copy, and install snippets to reflect the 8-language / 179-model PII catalog and uv pip install "openmed[hf]"

Fixed

Smart semantic-merge resolution no longer lets weaker model labels overwrite stronger validated pattern labels
Localized Dutch, Hindi, and Telugu month-name parsing now falls back correctly during date shifting instead of relying only on dateutil
Dutch phone, BSN, and street-address patterns were tightened after live smoke review to reduce overlap and improve entity labeling

Documentation

Updated README, REST service docs, website copy, notebook index, and the multilingual PII notebook for the v0.6.2 release

0.6.17 Mar 2026

OpenMed v0.6.1 introduces a Dockerized REST API MVP for deploying text analysis and PII workflows over HTTP.

OpenMed v0.6.1

OpenMed v0.6.1 introduces a Dockerized REST API MVP for deploying text analysis and PII workflows over HTTP.

Added

FastAPI service under openmed.service
GET /health for service status, version, and active profile
POST /analyze mapped to analyze_text(..., output_format="dict")
POST /pii/extract for multilingual PII extraction
POST /pii/deidentify for de-identification workflows
Request validation schemas and unit tests for API success, validation, and failure paths
CPU-focused Dockerfile and .dockerignore
REST service documentation and MkDocs navigation entry

Changed

Added optional service dependencies with fastapi and uvicorn[standard]
Expanded dev dependencies with fastapi and httpx for API testing
Updated the README, docs, and website content for REST deployment and multilingual PII support
Bumped the package version to 0.6.1

Notes

This is an MVP service layer intended to expose existing OpenMed Python APIs over HTTP.
Profile selection is supported through OPENMED_PROFILE.

What's Changed

Feature/v0.6.1 services by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/28

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.6.0...v0.6.1

Open source →

Additional notes

Added

Dockerized REST MVP for OpenMed service use-cases
- New FastAPI service module at openmed.service
- GET /health endpoint for service status and active profile reporting
- POST /analyze endpoint mapped to analyze_text(..., output_format="dict")
- POST /pii/extract endpoint mapped to extract_pii(...)
- POST /pii/deidentify endpoint mapped to deidentify(...)
Container runtime support
- New CPU-focused Dockerfile for service deployment
- Added .dockerignore for smaller build contexts
Service validation tests
- New unit tests covering endpoint success/failure paths, schema validation, and profile selection

Changed

Added optional service dependency extra in pyproject.toml (fastapi, uvicorn[standard])
Expanded dev extra with API test dependencies (fastapi, httpx)

Documentation

Added REST service guide: docs/rest-service.md
Added MkDocs navigation entry for REST service docs
Updated README with REST API and Docker usage examples

Open source →

0.6.028 Feb 2026

OpenMed v0.6.0 is a breaking, API-first release. This version removes terminal interfaces and streamlines the package around Python APIs for services,

OpenMed v0.6.0 (2026-02-28)

OpenMed v0.6.0 is a breaking, API-first release. This version removes terminal interfaces and streamlines the package around Python APIs for services, notebooks, and production pipelines.

Breaking changes

Removed the openmed console entrypoint from package metadata.
Removed CLI modules:
- openmed.cli
- openmed.zero_shot.cli
Removed TUI modules:
- openmed.tui
Removed cli_main from the top-level public API.
Removed optional dependency groups: cli, tui.

What changed

Documentation and website examples now focus on API usage (analyze_text, BatchProcessor, PII APIs).
MkDocs navigation removed CLI/TUI pages (docs/cli.md, docs/tui.md).
Release automation moved to tag-driven publishing:
- .github/workflows/publish.yml now triggers on push tags (v*) and validates tag/version alignment.
Version bumping now uses openmed/__about__.py as the source of truth.
- scripts/release/release.py now bumps version only.
Build/release commands updated to consistently use python3 in Make/release scripts.
Lockfile/dependencies cleaned up after CLI/TUI removal (including Typer/Textual/Rich-related packages).

Migration guide

If you were using CLI commands such as openmed analyze, openmed batch, openmed models, or openmed tui, migrate to Python APIs.

Analyze text

from openmed import analyze_text

result = analyze_text(
    "Patient diagnosed with chronic myeloid leukemia and started imatinib.",
    model_name="disease_detection_superclinical",
)
print(result.entities)

Batch processing

from openmed import BatchProcessor

processor = BatchProcessor(model_name="disease_detection_superclinical")
output = processor.process_texts([
    "Text one",
    "Text two",
])
print(output.summary())

What's Changed

v0.6.0 api only next release by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/26

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.8...v0.6.0

Open source →

Additional notes

Removed

CLI and TUI surfaces removed: OpenMed is now a Python API-first package
- Removed openmed console entrypoint from package metadata
- Removed openmed.cli and openmed.tui modules
- Removed zero-shot CLI modules under openmed.zero_shot.cli
- Removed cli_main from the top-level openmed public API

Changed

Updated package metadata to remove CLI/TUI extras (cli, tui)
Updated docs and website content to API-only guidance
Consolidated PyPI publishing into a single tag-driven workflow (publish.yml)
Updated release tooling to use openmed/__about__.py as the version source of truth

Open source →

0.5.819 Feb 2026

Feature/v0.5.8 fixes pii extraction by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/25

patch release

What's Changed

Feature/v0.5.8 fixes pii extraction by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/25

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.7...v0.5.8

Open source →

Additional notes

Fixed

PII replace label mapping coverage:
- Added robust normalization map so replacement data is generated for label variants (first_name, last_name, dob, postal_code, etc.)
- Expanded locale fake-data dictionaries with FIRST_NAME, LAST_NAME, and ZIPCODE values across supported languages
Span alignment stability:
- extract_pii() and deidentify() now strip leading/trailing whitespace before inference so spans remain aligned with analyze_text() validation behavior
Spanish accent remapping robustness:
- Added regression coverage for off-by-one spans combined with accent restoration

Open source →

0.5.718 Feb 2026

Fix: Added _fix_entity_spans() in OutputFormatter that extends end forward while the next character is still alphanumeric. This runs before entity gro

patch release

Fix: Added _fix_entity_spans() in OutputFormatter that extends end forward while the next character is still alphanumeric. This runs before entity grouping, so all downstream consumers (extraction, de-identification, smart merging) get correct spans.

What's Changed

Prepare v0.5.7 span-fix and Spanish notebook updates by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/24

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.6...v0.5.7

Open source →

Additional notes

Fixed

Entity span repair in output formatter:
- Added _fix_entity_spans() to correct tokenizer end-offset truncation and trim whitespace around predicted spans
- Integrated span repair into output formatting before grouping
Regression coverage:
- Added dedicated tests for off-by-one span fixes, whitespace trimming, and boundary handling
Documentation notebook refresh:
- Updated multilingual PII notebook examples to reflect span-fix behavior

Open source →

0.5.618 Feb 2026

This release adds full Spanish PII support and expands multilingual model coverage.

OpenMed v0.5.6

This release adds full Spanish PII support and expands multilingual model coverage.

Highlights

Added Spanish PII extraction and de-identification support via lang="es" in:
- extract_pii()
- deidentify()
Added Spanish national ID validators with checksum verification:
- validate_spanish_dni()
- validate_spanish_nie()
Added Spanish-specific detection patterns for:
- Dates (DD/MM/YYYY and 15 de enero de 2020 format)
- Phone numbers (+34, mobile/landline formats)
- National IDs (DNI/NIE)
- Street addresses
- Postal codes

Model Registry Updates

Added 2 new English base PII architectures:
- pii_biomed_bert_full (BiomedBERTFull-Base-110M)
- pii_lite_clinical_u (LiteClinicalU-Small-66M)
Added Spanish language model generation in registry (Spanish- prefix).
Expanded multilingual PII coverage to 176+ models total.
get_pii_models_by_language("es") now returns full Spanish model set.
get_default_pii_model("es") now returns the Spanish default model.

De-identification Improvements

Added accent normalization support for model inference:
- New normalize_accents parameter in extract_pii() and deidentify()
- Auto-enabled for Spanish by default
- Preserves entity offsets and original accented output text
Extended locale-aware date shifting/formatting to support Spanish conventions.

Locale Data

Added Spanish fake replacement data for replace workflows:
- Names, emails, phones, addresses, IDs, dates, locations

Testing

Added and updated unit tests for:
- Spanish DNI/NIE validation
- Spanish patterns and fake data
- Spanish model registry completeness and naming
- Accent normalization behavior
Test run for updated suites: 170 passed

Compatibility

No breaking API changes intended.
Version updated to 0.5.6.

What's Changed

Add Spanish PII support and cut v0.5.6 by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/23

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.5...v0.5.6

Open source →

Additional notes

Added

Spanish PII Detection & De-identification: Full Spanish language support for PII extraction
- extract_pii() and deidentify() now accept lang="es" for Spanish clinical text
- Automatic model selection for Spanish — correct language-specific model chosen when lang="es"
- 7 new Spanish-specific regex patterns for dates, phone numbers, addresses, postal codes, and national IDs
- Spanish date format support with unique "de" connector (e.g., "15 de enero de 2020")
Spanish National ID Validators: DNI and NIE document validation with checksum verification
- validate_spanish_dni() — Spanish DNI 8-digit + check letter (mod-23 lookup table)
- validate_spanish_nie() — Spanish NIE with X/Y/Z prefix conversion and DNI algorithm
2 New English Base Model Architectures: Expanded PII model coverage
- pii_biomed_bert_full — BiomedBERTFull-Base-110M for comprehensive biomedical PII detection
- pii_lite_clinical_u — LiteClinicalU-Small-66M for universal lightweight PII detection
- Both architectures auto-generate variants for all 5 supported languages
Expanded Model Registry: 35 Spanish PII models + 8 new models across existing languages
- Total PII models expanded from 133 to 176+ (36 English + 35 x 4 languages)
- get_pii_models_by_language("es") returns all 35 Spanish models
- get_default_pii_model("es") returns the recommended Spanish default model
Accent Normalization: Transparent accent stripping for models trained on accent-free text
- normalize_accents parameter on extract_pii() and deidentify() (auto-enabled for Spanish)
- Strips diacritical marks before model inference, maps entity positions back to original accented text
- _strip_accents() helper preserves character count via NFC/NFD normalization
- Can be explicitly enabled (normalize_accents=True) or disabled (normalize_accents=False) for any language
Spanish Locale Data: Culturally appropriate synthetic data for the replace method
- Spanish fake names, emails, phone numbers (+34), addresses, and IDs (DNI/NIE)
- Spanish month names for date parsing and formatting
- European DD/MM/YYYY date handling for Spanish
Testing: Comprehensive Spanish PII test coverage
- Spanish DNI validator tests (6 tests) and NIE validator tests (6 tests)
- Spanish pattern matching tests for dates, phones, DNI, NIE
- Spanish model registry tests: count, naming, mirror structure
- Updated existing tests: fixed "es" to "ja" in unsupported language assertions

Changed

_LANGUAGE_CONFIG in model registry now includes "es": {"name": "Spanish", "prefix": "Spanish-"}
French, German, and Italian model counts updated from 33 to 35 per language (2 new base architectures)
SUPPORTED_LANGUAGES expanded to include "es"
Date handling functions (_shift_date, _shift_date_basic, _format_date_like_original) now support Spanish

0.5.511 Feb 2026

OpenMed — v0.5.5: Multilingual PII Detection & De-identification

Highlights

OpenMed now supports multilingual PII detection and de-identification for clinical text in English, French, German, and Italian. The model catalog has expanded from 33 to 132+ PII models, and all core APIs remain fully backward compatible.

New Features

Multilingual PII Extraction & De-identification

extract_pii() and deidentify() now accept a lang parameter (ISO 639-1: en, fr, de, it) for language-aware processing
Automatic model selection — the correct language-specific model is chosen when lang is specified
Language-specific regex patterns for dates, phone numbers, addresses, postal codes, and national IDs

National ID Validators

French NIR/INSEE (15-digit social security numbers with mod-97 checksum)
German Steuer-ID (11-digit tax identification numbers)
Italian Codice Fiscale (16-character alphanumeric fiscal codes)

Locale-Aware Date Handling

European day-first parsing for fr/de/it (DD/MM/YYYY, DD.MM.YYYY)
US month-first parsing for en (MM/DD/YYYY)
Localized month names preserved during date shifting

Culturally Appropriate De-identification

Language-specific fake data (names, emails, phone numbers, addresses, etc.) used by the replace method
Ensures de-identified text remains natural and realistic per locale

Expanded Model Registry

~99 new multilingual PII models generated across 33 architectures x 3 new languages
New APIs: get_pii_models_by_language(), get_default_pii_model()

New Module

openmed/core/pii_i18n.py — Internationalization module with language-specific patterns, validators, fake data, and month names

New Exports

SUPPORTED_LANGUAGES, DEFAULT_PII_MODELS, LANGUAGE_PII_PATTERNS, get_patterns_for_language
get_pii_models_by_language, get_default_pii_model

Improvements

Entity Merger — Label normalization now handles cross-language entity types (nir, insee, steuer_id, codice_fiscale, postcode variants)
Specificity Hierarchy — Updated to correctly prioritize language-specific entity sub-types during merging

Documentation

New tutorial notebook: Multilingual PII Detection Guide (examples/notebooks/Multilingual_PII_Detection_Guide.ipynb) covering all 4 languages, cross-language comparison, batch processing, and custom model selection

Tests

test_pii_i18n.py — 373 lines of unit tests for the new i18n module
test_model_registry_multilingual.py — 202 lines of unit tests for multilingual model generation
Updated test_pii.py and test_pii_entity_merger.py with multilingual test cases

Backward Compatibility

Fully backward compatible. All existing English-only usage works unchanged without any code modifications.

What's Changed

Multilingual PII Detection & De-identification by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/21

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.1...v0.5.5

Open source →

Additional notes

Added

Multilingual PII Detection & De-identification: Language-aware PII extraction for clinical text
- extract_pii() and deidentify() now accept a lang parameter (ISO 639-1: en, fr, de, it)
- Automatic model selection — correct language-specific model chosen when lang is specified
- Language-specific regex patterns for dates, phone numbers, addresses, postal codes, and national IDs
- 18 new regex patterns (6 per language) for French, German, and Italian
National ID Validators: Country-specific document validation with checksum verification
- validate_french_nir() — French NIR/INSEE 15-digit social security numbers (mod-97 checksum)
- validate_german_steuer_id() — German 11-digit tax identification numbers (digit-frequency rules)
- validate_italian_codice_fiscale() — Italian 16-character alphanumeric fiscal codes
Locale-Aware Date Handling: Language-appropriate date parsing and formatting
- European day-first parsing for fr/de/it (DD/MM/YYYY, DD.MM.YYYY)
- US month-first parsing for en (MM/DD/YYYY)
- Localized month names preserved during date shifting
Culturally Appropriate De-identification: Language-specific synthetic data for the replace method
- Fake names, emails, phone numbers, addresses, and IDs per locale
- LANGUAGE_FAKE_DATA dictionary for English, French, German, and Italian
Expanded Model Registry: Multilingual model generation across all PII architectures
- ~99 new multilingual PII models (33 architectures x 3 new languages)
- Total PII models expanded from 33 to 132+
- get_pii_models_by_language() — returns all PII models for a given language
- get_default_pii_model() — returns the recommended default model for a language
New Module: openmed/core/pii_i18n.py — Internationalization module
- SUPPORTED_LANGUAGES, DEFAULT_PII_MODELS, LANGUAGE_PII_PATTERNS constants
- get_patterns_for_language() — returns combined English + language-specific regex patterns
- LANGUAGE_MONTH_NAMES dictionary with month names in all 4 languages
Documentation
- New Multilingual PII Detection Guide notebook
  - Cross-language comparison, batch processing, and custom model selection
  - Examples for French, German, and Italian clinical notes
  - All de-identification methods with multilingual fake data
Testing
- test_pii_i18n.py — unit tests for the i18n module (373 lines)
- test_model_registry_multilingual.py — unit tests for multilingual model generation (202 lines)
- Updated test_pii.py and test_pii_entity_merger.py with multilingual test cases

Changed

_redact_entity() and _generate_fake_pii() now propagate lang parameter for language-appropriate replacements
normalize_label() handles national ID variants (nir, insee, steuer_id, codice_fiscale) and postal code variants (postcode, zipcode, postal_code)
Label specificity hierarchy expanded with national_id sub-types for cross-language entity resolution
CATEGORIES["Privacy"] dynamically includes all PII model keys (English + multilingual)
Updated __init__.py exports with multilingual PII support functions

0.5.114 Jan 20262 sources agree

Context-Aware PII Scoring: Presidio-inspired confidence scoring system

Added

Context-Aware PII Scoring: Presidio-inspired confidence scoring system
- PIIPattern dataclass extended with base_score, context_words, context_boost, and validator fields
- Context detection via find_context_words() - boosts confidence when keywords like "SSN:", "DOB:", "NPI:" appear near detected entities
- Checksum validation functions: validate_ssn(), validate_luhn() (credit cards), validate_npi(), validate_phone_us()
- Invalid matches (e.g., SSN starting with 000 or 666) get reduced confidence scores
- Combined model + pattern scoring (60/40 weighted average) for optimal accuracy
- Low base scores prevent false positives; context words confirm true PHI
Website Updates
- New "Clinical Text De-Identification" section on landing page
- Key stats row: 18+ PHI types, 100% local processing, $0 API fees, Apache-2.0
- Six feature cards: Context-Aware Detection, Checksum Validation, Smart Merging, Zero Data Movement, Flexible Redaction, HIPAA Safe Harbor
- Syntax-highlighted code example with correct API usage
- CTA buttons linking to documentation and HuggingFace models

Changed

Updated default PII detection model name to OpenMed-PII-SuperClinical-Small-44M-v1
merge_entities_with_semantic_units() now supports context-aware pattern scoring

Fixed

MkDocs navigation: Added medical-tokenizer.md and pii-smart-merging.md to nav structure
Broken link in cli.md to PII notebook (now links to GitHub)
Broken links in pii-smart-merging.md to non-existent documentation pages
Website code example now uses correct API (entity.text, entity.label, entity.confidence)

What's Changed

Feature/v0.5.0 pii extraction by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/19

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.5.0...v0.5.1

Open source →

0.5.014 Jan 20262 sources agree

PII Detection & De-identification: HIPAA-compliant PII extraction and de-identification

Added

PII Detection & De-identification: HIPAA-compliant PII extraction and de-identification
- extract_pii() function for detecting PII entities in clinical text
- deidentify() function with 5 de-identification methods:
  - mask: Replace with placeholders ([NAME], [DATE], etc.)
  - remove: Complete removal of PII entities
  - replace: Replace with synthetic data
  - hash: Cryptographic hashing for record linking
  - shift_dates: Shift dates while preserving temporal relationships
- reidentify() function for reversing de-identification with stored mappings
- Support for all 18 HIPAA Safe Harbor identifiers
- Configurable confidence thresholds for precision/recall control
- Batch processing support for PII extraction and de-identification
- PIIEntity and DeidentificationResult dataclasses
Smart Entity Merging: Advanced post-processing to fix tokenization fragmentation
- Regex-based semantic unit detection with 20+ PII patterns
- Automatic merging of fragmented entities (e.g., dates split as "01" + "/15/1970" → "01/15/1970")
- Dominant label selection with confidence-based tie-breaking
- Label specificity hierarchy (e.g., date_of_birth > date)
- Support for dates (6 formats), SSN, phone numbers, emails, URLs, addresses, IP addresses, MAC addresses, ZIPs, credit cards
- Custom pattern support via PIIPattern class
- Enabled by default with use_smart_merging=True parameter
- Public API exports: merge_entities_with_semantic_units(), find_semantic_units(), calculate_dominant_label(), PII_PATTERNS
- Minimal performance overhead (~5-10%)
PII CLI Commands: Comprehensive command-line interface for PII operations
- openmed pii extract: Extract PII entities from text or files
- openmed pii deidentify: De-identify text or files with method selection
- openmed pii batch-extract: Batch PII extraction from directories
- openmed pii batch-deidentify: Batch de-identification with method selection
- All commands support confidence thresholds, smart merging, and output formatting
- Date shifting parameter (--date-shift-days) for temporal preservation
PII TUI Mode: Interactive PII detection in the terminal interface
- Visual PII entity highlighting with color coding
- Real-time de-identification preview
- Model selection for PII detection models
PII Model Registry: Added PII detection models
- pii_detection_superclinical (434M parameters)
- Covers 18+ PII entity types (names, dates, SSN, phone, email, addresses, medical records, etc.)
Comprehensive Documentation
- PII Detection & Smart Merging Guide (452 lines)
  - Algorithm explanation and implementation details
  - Complete API reference with examples
  - Supported PII patterns catalog
  - Performance characteristics
  - Troubleshooting guide
- Complete PII Jupyter Notebook (48 cells)
  - Step-by-step tutorial covering all PII functionality
  - Before/after smart merging comparisons
  - All 5 de-identification methods demonstrated
  - Re-identification workflows
  - Batch processing examples
  - Confidence thresholding guidelines
  - Custom PII patterns
  - Clinical use cases (discharge summaries, research datasets, HIPAA compliance)
  - HTML visualization examples
  - CLI usage reference
  - Best practices and security considerations
- Notebooks README
  - Navigation guide for all notebooks
  - Learning paths for different user types
  - Quick reference table
- Updated README.md with PII capabilities
- Updated CLI documentation with PII commands
- Updated feature map and documentation index
Testing
- Comprehensive PII extraction and de-identification test suite
- Smart entity merging validation tests
- All 5 de-identification methods tested
- Complex clinical note integration tests

Changed

Default PII extraction behavior now uses smart entity merging (use_smart_merging=True)
Enhanced model registry with PII detection category

What's Changed

Feature/v0.5.0 pii extraction by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/18

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.4.0...v0.5.0

Open source →

0.4.029 Dec 20252 sources agree

Interactive TUI (Terminal User Interface): Full-featured terminal workbench for clinical NER analysis

Added

Interactive TUI (Terminal User Interface): Full-featured terminal workbench for clinical NER analysis
- Rich text input with multi-line support
- Color-coded entity highlighting in annotated view
- Entity table with confidence bars sorted by score
- Model switcher modal (F2) for switching between models
- Configuration panel (F3) for adjusting threshold and settings
- Profile switcher (F4) for quick dev/prod/test/fast presets
- Analysis history (F5) with recall and deletion
- Export results (F6) to JSON, CSV, or clipboard
- File navigation (Ctrl+O) for loading text files
- Status bar showing model, profile, threshold, and inference time
- CLI command: openmed or openmed tui
TUI Documentation: Comprehensive guide at docs/tui.md
- Interface overview with ASCII preview
- Keyboard shortcuts reference
- Profile presets documentation
- Export format examples
- Python API usage
Website Updates
- New Python Toolkit section showcasing TUI, CLI, batch processing, and profiles
- Interactive TUI preview with color-coded entities
- CLI and TUI tabs in hero code block
- Updated software version metadata

Changed

Updated mkdocs navigation to include TUI documentation

What's Changed

Feature/v0.4.0 tui by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/17

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.3.0...v0.4.0

Open source →

0.3.026 Dec 20252 sources agree

Batch Processing: Process multiple texts or files in a single operation

Added

Batch Processing: Process multiple texts or files in a single operation
- BatchProcessor class for configurable batch operations
- BatchItem, BatchItemResult, BatchResult dataclasses
- process_batch() convenience function
- File discovery with glob patterns and recursive search
- Progress callbacks for monitoring long-running jobs
- Configurable error handling (fail-fast or continue)
- CLI batch command with full feature support
Configuration Profiles: Named configuration presets for different environments
- Built-in profiles: dev, prod, test, fast
- OpenMedConfig.from_profile() and with_profile() methods
- list_profiles(), get_profile(), save_profile(), delete_profile() functions
- Custom profile persistence to disk
- CLI commands: config profiles, profile-show, profile-use, profile-save, profile-delete
- --profile flag for config show command
Performance Profiling: Built-in timing and metrics utilities
- Timer context manager for measuring code blocks
- Profiler class for tracking metrics across multiple runs
- @profile decorator for easy function profiling
- ProfilingMetrics dataclass for structured timing data
- Support for nested profiling and statistical aggregation
Documentation
- New Batch Processing guide
- New Configuration Profiles guide
- New Performance Profiling guide
- Updated CLI documentation with new commands
- Updated feature map and documentation index
Testing
- 89 new unit tests for batch, profiles, and profiling modules
- Total test count: 218 passing tests

Open source →

0.2.220 Dec 2025

Medical-aware tokenization is now remap-only: the model always uses its own tokenizer/vocab; we tokenize separately for output and remap spans to clea

Highlights

Medical-aware tokenization is now remap-only: the model always uses its own tokenizer/vocab; we tokenize separately for output and remap spans to cleaner clinical tokens (e.g., IL-6-mediated, CAR-T, t(8;21)).
New utilities: medical_tokenize and remap_predictions_to_tokens for span-level postprocessing.
CLI/config flags now control remapping (not model tokenization). Use --use-medical-tokenizer/--no-medical-tokenizer or config/env overrides.
Added comparison script examples/custom_tokenizer/compare_medical_remap.py to see outputs with remap on/off.
Documentation updated across README, CLI guide, analyze-text guide, and dedicated docs/medical-tokenizer.md.
Full test suite passing.

GLiNER2 groundwork

Added GLINER2 model family enum, indexing hints, availability checks, loader wrapper (openmed/ner/families/gliner2.py), and inference path in ner.infer.
Public exports include ensure_gliner2_available, is_gliner2_available, and load_gliner2_handle.

CLI extras

New optional cli extra (Typer/Rich) and draft Typer-based CLI at openmed/cli/typer_app.py.

Bug/behavior fixes

Removed prior behavior that injected medical exceptions into the model tokenizer/vocab (which could distort embeddings).
Integration test updated to reflect remap-disabled legacy expectations.

Examples & notebooks

New comparison script plus existing demos under examples/custom_tokenizer/.
Notebook benchmark examples/notebooks/Medical_Tokenizer_Benchmark.ipynb demonstrates remap on/off latency and entities.

How to use

Default: remap enabled. Disable with OPENMED_USE_MEDICAL_TOKENIZER=0 or --no-medical-tokenizer.
Add protected terms: OPENMED_MEDICAL_TOKENIZER_EXCEPTIONS="MY-DRUG-123,ABC-001" or CLI flag --medical-tokenizer-exceptions.

Open source →

Additional notes

Added

Medical-aware tokenizer with customizable exceptions
CLI --use-medical-tokenizer and --medical-tokenizer-exceptions flags

Fixed

Token boundary issues with medical terminology

Open source →

0.2.110 Dec 2025

Medical-aware pre-tokenizer now ships in-core and is on by default for fast HF tokenizers; it uses Bert-style splitting plus a curated list of biomedi

Custom Tokenizer

Medical-aware pre-tokenizer now ships in-core and is on by default for fast HF tokenizers; it uses Bert-style splitting plus a curated list of biomedical exceptions (COVID-19, IL-6, CAR-T, etc.) and injects user-provided exceptions. Falls back silently if tokenizers isn’t installed or the tokenizer is slow.

DEFAULT_MEDICAL_EXCEPTIONS = [
    "COVID-19",
    "SARS-CoV-2",
    "IL-6",
    "IL-2",
    "TNF-alpha",
    "BCR-ABL1",
    "CAR-T",
    "post-CAR-T",
    "t(8;21)",
    "t(15;17)",
]
def build_medical_pretokenizer(exceptions: Optional[Iterable[str]] = None):
    ...
def apply_medical_pretokenizer(tokenizer: Any, exceptions: Optional[Iterable[str]] = None) -> bool:
    ...

Config/CLI: OpenMedConfig gains use_medical_tokenizer (default True) and medical_tokenizer_exceptions; CLI adds --use-medical-tokenizer/--no-medical-tokenizer and --medical-tokenizer-exceptions, with env overrides (OPENMED_USE_MEDICAL_TOKENIZER, OPENMED_MEDICAL_TOKENIZER_EXCEPTIONS) honored.

# Medical-aware pre-tokenizer toggle (fast tokenizers only)
use_medical_tokenizer: bool = True
# Optional list of hyphenated/clinical exceptions to keep intact
medical_tokenizer_exceptions: Optional[List[str]] = None

analyze_parser.add_argument(
    "--use-medical-tokenizer",
    dest="use_medical_tokenizer",
    action="store_true",
    default=None,
    help="Force-enable the medical-aware pre-tokenizer (default from config).",
)
analyze_parser.add_argument(
    "--no-medical-tokenizer",
    dest="use_medical_tokenizer",
    action="store_false",
    default=None,
    help="Disable the medical-aware pre-tokenizer and fall back to the model default.",
)
analyze_parser.add_argument(
    "--medical-tokenizer-exceptions",
    default=None,
    help="Comma-separated extra terms to keep intact (e.g., MY-DRUG-123,ABC-001).",
)

Model loading/pipelines automatically apply the medical pre-tokenizer when enabled, including passing custom exceptions into both direct loads and pipelines.

if getattr(self.config, "use_medical_tokenizer", False):
    ...
    applied = apply_medical_pretokenizer(
        tokenizer,
        exceptions=getattr(self.config, "medical_tokenizer_exceptions", None)
    )
...
if getattr(self.config, "use_medical_tokenizer", False):
    ...
    applied = apply_medical_pretokenizer(
        ner_pipeline.tokenizer,
        exceptions=getattr(self.config, "medical_tokenizer_exceptions", None),
    )

Tokenization utilities add infer_tokenizer_max_length and helper alignment/truncation utilities; CLI/registry introspection can surface sensible max lengths.
Docs/examples: new docs/medical-tokenizer.md, README/CLI docs updates, runnable examples in examples/custom_tokenizer/ plus benchmark/demo notebooks.
Tests: coverage for medical pre-tokenizer behavior and CLI config overrides (tests/test_medical_tokenizer.py, tests/test_cli_medical_tokenizer.py).
Dependency: tokenizers>=0.15 added to the hf extra in pyproject.toml; enable the medical tokenizer by installing with [hf].

Notable behavior change: fast tokenizers now default to the medical pre-tokenizer; disable via config/CLI/env if you need raw tokenizer parity with prior runs or baseline comparisons.

What's Changed

implementing Custom tokenizer by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/13
Update about.py by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/14

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.2.0...v0.2.1

Open source →

Additional notes

Added

GLiNER2 support for zero-shot NER
Enhanced model registry with GLiNER2 family

Open source →

0.2.029 Oct 2025

Sentence detection and segmentation

New features
- Sentence detection and segmentation
  - Added pySBD-based sentence segmentation utilities (openmed.processing.sentences).
  - Introduced SentenceSpan dataclass and segment_text(text, language, clean, segmenter).
- Sentence-aware inference in analyze_text
  - New options: sentence_detection (default: True), sentence_language, sentence_clean, sentence_segmenter.
  - analyze_text now segments text, trims whitespace, suppresses placeholder-only segments, groups sentences into chunks, and batches chunked inputs to the NER pipeline.
  - Predictions are mapped back to original document coordinates and annotated with sentence metadata (sentence_index, sentence_text, sentence_start, sentence_end).
- Output metadata
  - EntityPrediction now contains metadata.
  - JSON/CSV output includes sentence_index and sentence_text when available; other metadata fields are exported with metadata_<key>.
- CLI improvement
  - CLI lazy-loads heavy APIs (analyze_text, get_model_max_length, list_models) to reduce startup cost and avoid importing large deps at CLI parse time.
- Introduces the “zeroshot” development changes — code and assets to support zero-shot workflows or inference (new features and helpers added).
- Adds significant implementation code across multiple modules (bulk of the +2.4k LOC).
- Small, targeted deletions (10 LOC) to remove/replace previous behavior.
Improvements
- Better tokenizer handling
  - Respect max_length vs truncation: when truncation is enabled the tokenizer.model_max_length is set to the desired max; when truncation is disabled the tokenizer.model_max_length is set to 0.
- Robust pipeline normalization & batching
  - Normalizes varied pipeline return shapes, flattens and adjusts predictions from chunked/batched runs.
- Grouping / merging refinement
  - Grouping/merging of adjacent entities now respects sentence boundaries (won’t merge entities across sentences).
  - Merged entities carry metadata from the first entity when present.
- Error handling
  - Hugging Face import checks expanded to catch OSError in addition to ImportError (more robust on some platforms).
- Public API exports
  - processing.init now exposes the sentences module.
Documentation & packaging
- README updates
  - Clarified HF token wording and optional extras; added "Install with uv" usage and examples for installing the hf extra and dev extras.
  - Example PyTorch install pinned to torch==2.9.0 (CPU example) and note about swapping index URL for CUDA.
- pyproject.toml
  - Added pysbd >=0.3.4,<0.4 to project dependencies.
- Includes multiple commits that appear to add or update tests and/or documentation accompanying the new features.
- Code refactors and reorganizations to support the new zeroshot functionality.
Tests & fixtures
- New tests:
  - Integration tests exercising sentence detection with real models.
  - Unit tests updated/added to cover sentence-detection behavior, truncation handling, batching, and metadata attachment.
- New fixtures:
  - tests/fixtures/clinical_note.txt
  - tests/fixtures/long_clinical_note.txt
Upgrade / compatibility notes
- analyze_text default behavior: sentence_detection is enabled by default. If you want pre-0.2.0 behavior, call analyze_text(..., sentence_detection=False).
- pysbd: sentence detection requires pysbd; the code raises ImportError when pySBD is not available (and will disable sentence detection). The project declares a pysbd dependency.
- Consumers of programmatic outputs should expect metadata fields on EntityPrediction (and possible new columns in CSV/keys in JSON).
- CLI now lazy-loads heavy modules; no functional CLI change but improved startup characteristics.
- Several files updated to integrate the new feature into the codebase (likely API/CLI hooks, configuration or example notebooks).
- No large dependency removals/additions noted in the PR metadata.

What's Changed

Add fast sentence detection by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/6
Update version to 0.2.0-rc1 in about.py to reflect the release ca… by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/7
Add notebook for sentence-level batching with analyze_text: Demonstra… by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/8
Dev zeroshot by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/9
Update version to 0.2.0-rc2 in about.py: Increment version number… by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/10
Dev zeroshot by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/11
Docs by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/12

New Contributors

@maziyarpanahi made their first contribution in https://github.com/maziyarpanahi/openmed/pull/6

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.10...v0.2.0

Open source →

Additional notes

Added

Typer-based CLI interface (openmed command)
analyze command for single text analysis
models list and models info commands
config show and config set commands
Rich terminal output formatting

Changed

Migrated CLI from argparse to Typer

Open source →

0.2.0rc26 Nov 2025pre-release

New / changed functionality (high-level)

New / changed functionality (high-level)
- Introduces the “zeroshot” development changes — code and assets to support zero-shot workflows or inference (new features and helpers added).
- Adds significant implementation code across multiple modules (bulk of the +2.4k LOC).
- Small, targeted deletions (10 LOC) to remove/replace previous behavior.
Quality, docs & tests
- Includes multiple commits that appear to add or update tests and/or documentation accompanying the new features.
- Code refactors and reorganizations to support the new zeroshot functionality.
Maintenance & developer experience
- Several files updated to integrate the new feature into the codebase (likely API/CLI hooks, configuration or example notebooks).
- No large dependency removals/additions noted in the PR metadata.

What's Changed

Add notebook for sentence-level batching with analyze_text: Demonstra… by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/8
Dev zeroshot by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/9
Update version to 0.2.0-rc2 in about.py: Increment version number… by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/10

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.2.0-rc1...v0.2.0-rc2

Open source →

0.2.0rc129 Oct 2025pre-release

Sentence detection and segmentation

New features
- Sentence detection and segmentation
  - Added pySBD-based sentence segmentation utilities (openmed.processing.sentences).
  - Introduced SentenceSpan dataclass and segment_text(text, language, clean, segmenter).
- Sentence-aware inference in analyze_text
  - New options: sentence_detection (default: True), sentence_language, sentence_clean, sentence_segmenter.
  - analyze_text now segments text, trims whitespace, suppresses placeholder-only segments, groups sentences into chunks, and batches chunked inputs to the NER pipeline.
  - Predictions are mapped back to original document coordinates and annotated with sentence metadata (sentence_index, sentence_text, sentence_start, sentence_end).
- Output metadata
  - EntityPrediction now contains metadata.
  - JSON/CSV output includes sentence_index and sentence_text when available; other metadata fields are exported with metadata_<key>.
- CLI improvement
  - CLI lazy-loads heavy APIs (analyze_text, get_model_max_length, list_models) to reduce startup cost and avoid importing large deps at CLI parse time.
Improvements
- Better tokenizer handling
  - Respect max_length vs truncation: when truncation is enabled the tokenizer.model_max_length is set to the desired max; when truncation is disabled the tokenizer.model_max_length is set to 0.
- Robust pipeline normalization & batching
  - Normalizes varied pipeline return shapes, flattens and adjusts predictions from chunked/batched runs.
- Grouping / merging refinement
  - Grouping/merging of adjacent entities now respects sentence boundaries (won’t merge entities across sentences).
  - Merged entities carry metadata from the first entity when present.
- Error handling
  - Hugging Face import checks expanded to catch OSError in addition to ImportError (more robust on some platforms).
- Public API exports
  - processing.init now exposes the sentences module.
Documentation & packaging
- README updates
  - Clarified HF token wording and optional extras; added "Install with uv" usage and examples for installing the hf extra and dev extras.
  - Example PyTorch install pinned to torch==2.9.0 (CPU example) and note about swapping index URL for CUDA.
- pyproject.toml
  - Added pysbd >=0.3.4,<0.4 to project dependencies.
Tests & fixtures
- New tests:
  - Integration tests exercising sentence detection with real models.
  - Unit tests updated/added to cover sentence-detection behavior, truncation handling, batching, and metadata attachment.
- New fixtures:
  - tests/fixtures/clinical_note.txt
  - tests/fixtures/long_clinical_note.txt
Upgrade / compatibility notes
- analyze_text default behavior: sentence_detection is enabled by default. If you want pre-0.2.0 behavior, call analyze_text(..., sentence_detection=False).
- pysbd: sentence detection requires pysbd; the code raises ImportError when pySBD is not available (and will disable sentence detection). The project declares a pysbd dependency.
- Consumers of programmatic outputs should expect metadata fields on EntityPrediction (and possible new columns in CSV/keys in JSON).
- CLI now lazy-loads heavy modules; no functional CLI change but improved startup characteristics.

What's Changed

Add fast sentence detection by @maziyarpanahi in https://github.com/maziyarpanahi/openmed/pull/6

New Contributors

@maziyarpanahi made their first contribution in https://github.com/maziyarpanahi/openmed/pull/6

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.10...v0.2.0-rc1

Open source →

0.1.1017 Oct 2025

Introduced the openmed command-line interface with analyze, models, and config subcommands, making it easy to run inference, discover models, and mana

Introduced the openmed command-line interface with analyze, models, and config subcommands, making it easy to run inference, discover models, and manage settings from the terminal (includes new console script wiring and entry points).
Extended CLI capabilities: models list now supports --include-remote Hugging Face discovery, models info surfaces inferred max sequence length, and configuration helpers honor XDG paths and environment overrides.
Added ModelLoader.get_max_sequence_length, the top-level get_model_max_length, and metadata propagation so analyze_text reports the effective context window while keeping tokenizer settings in sync.
Fixed a regression where transformers pipelines rejected the truncation kwarg, ensuring the CLI and library work with the updated transformers>=4.50 API.
Bumped Hugging Face dependencies to transformers>=4.50 / huggingface-hub>=0.30 and documented the Python ≥3.10 requirement for better upstream compatibility.
Added scripts/reset_uv_env_and_run_tests.sh to automate nuking the environment, recreating a UV-managed Python 3.11 venv, reinstalling OpenMed from source, and running the full (slow-inclusive) pytest suite.
Expanded automated coverage with new CLI unit tests, max-length assertions, and updates to the test runner so CLI behavior ships with regression protection.
Refreshed documentation and notebooks to showcase the CLI workflow, remote model discovery, and context-length inspection so users can follow along in guides and Colab.
Simplified release plumbing (Hatch/Twine workflows, keyring handling, version metadata) to streamline future publishes.

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.9...v0.1.10

Open source →

Additional notes

Added

Initial public release
Core NER pipeline with HuggingFace integration
Model registry with curated biomedical models
analyze_text() one-call inference API
Advanced NER post-processing (grouping, filtering)
Multiple output formats (dict, JSON, HTML, CSV)
YAML/ENV configuration via OpenMedConfig
Zero-shot toolkit with GLiNER support

Open source →

0.1.10rc117 Oct 2025pre-release

Nothing published for this version

0.1.917 Oct 2025

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.8...v0.1.9

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.8...v0.1.9

Open source →

0.1.9rc117 Oct 2025pre-release

Nothing published for this version

0.1.817 Oct 2025

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.7...v0.1.8

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.7...v0.1.8

Open source →

0.1.8rc217 Oct 2025pre-release

Nothing published for this version

0.1.8rc017 Oct 2025pre-release

Nothing published for this version

0.1.712 Oct 2025

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.6...v0.1.7

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v0.1.6...v0.1.7

Open source →

0.1.56 Oct 2025

Full Changelog: https://github.com/maziyarpanahi/openmed/commits/0.1.5

Full Changelog: https://github.com/maziyarpanahi/openmed/commits/0.1.5

Open source →

0.1.46 Oct 2025

Nothing published for this version

0.1.110 Sept 2025

Nothing published for this version

0.19 Aug 2025

Nothing published for this version

Your coding agent can read these notes before it upgrades. Set up the MCP server →

openmed

Release timeline

Releases

OpenMed 1.7.0 is the multimodal de-identification and evaluation-depth release.

Highlights

The Big Picture

Multimodal And Document Redaction

Interoperability And Structured Health Data

Clinical Extraction And Context

Language, Locale, And Identifier Coverage

Runtime, Service, And Clients

Evaluation, Risk, And Release Evidence

Models, Backends, And Training

Security, Supply Chain, And Docs

Fixes

Migration Notes

PR Review Summary

Open Items Before Tagging

Verification For This Draft

What's Changed

Added

Changed

Fixed

Security

Dependencies

Removed

Upgrade Notes

OpenMed 1.6.0 is the privacy assurance and release-governance release.

Highlights

The Big Picture

Privacy Pipeline

Audit And Risk

Evaluation And Release Gates

Clinical And Interoperability

Model Manifest And Publishing

CLI, Docs, And Examples

Security And CI

Fixes

Migration Notes

PR Review Summary

Open Items Before Tagging

Verification For This Draft

What's Changed

Added

Changed

Fixed

Security

Tests

Notes

OpenMed 1.5.5 is a local-first PII, service lifecycle, and README refresh release.

Highlights

What's Changed

What's Changed

New Contributors

Added

Changed

Fixed

OpenMed 1.5.2 is a security and MLX conversion hardening release.

Highlights

Security

What's Changed

New Contributors

Security

Changed

Fixed

Tests

Changed

OpenMed v1.5.0 with Arabic, Japanese, and Turkish support

Highlights

MLX Rollout

SDK Updates

Docs And Release Prep

Verification

What's Changed

Added

Changed

Fixed

Tests

OpenMed v1.4.0 is the multilingual Privacy Filter release.

Highlights

OpenMed `1.7.0` is the multimodal de-identification and evaluation-depth release.

OpenMed `1.6.0` is the privacy assurance and release-governance release.

OpenMed `1.5.5` is a local-first PII, service lifecycle, and README refresh release.

OpenMed `1.5.2` is a security and MLX conversion hardening release.