docling-ibm-models

Fix

• tableformer: Consolidate post-processing source fixes (#176, #159, #175) (#185) (b234bf1)

Performance

• tableformer: O(1) orphan pdf-cell lookup in recovery fallbacks (#186) (bfa5c3c)

Feature

• Remove unused models (#184) (c875daf)

Breaking

• remove unused models (#184) (c875daf)

Feature

• reading-order: Re-enable same-row left-to-right linking with additive r2l fix (#163) (080aded)

Fix

• Correct graphics-to-captions assignment in reading order (#183) (cfd1d08)
• deps: Exclude transformers 5.13.0, which breaks the image-processor registration (#178) (5354f0f)
• UnboundLocalError in predict_dummy when no bboxes are detected (#147) (6974dfe)
• tf_cell_matcher: Correct the vertical non-overlap check in find_intersection (#142) (#177) (a81c60a)
• list_item_normalizer: Keep multi-line list item text (#179) (#181) (36fe1d0)

Performance

• Cache per-instance logger in _log() to remove millions of get_custom_logger calls (#182) (68c83ea)

Feature

• pdf: Recognize compound and hierarchical list-item markers (#169) (5787142)

Fix

• reading-order: Keep sequence-interruption query rectangle well-formed (#174) (b865f71)

Fix

• Pin transformers<5.9.0 to solve MPS issue (#164) (1483008)

Documentation

• security: Document security processes (#160) (cddcfae)

Fix

• Load code formula model with weights_only=True (#158) (2ebbcad)

Fix

• Load TableFormer model with torch.load(..., weights_only=True) (#157) (3761073)
• Extend predict_merges patterns for soft hyphen and Unicode letters (#155) (c5e0d0f)
• Chain merge continuations instead of stopping after one pair (#156) (d041ecc)

Feature

• Transformers to v5 (#151) (4fa73b9)

Feature

• TableFormer v2 (#149) (b99c955)

Feature

• Drop python 3.9 (#144) (d27b0f1)

Fix

• Deterministic page sorting in reading-order model (#141) (15ff139)

Fix

• ReadingOrderPredictor: Make _predict_page threadsafe (#137) (78473fe)

Fix

• Replace the depth-first-search implementations in ReadingOrderPredictor with non-recursive ones (#135) (78720cb)

Feature

• Support for python 3.14 (#131) (e21d57c)

Fix

• Allow flexibility for opencv (#130) (b9ff7eb)

Fix

• ReadingOrderPredictor: Improve the algorithm for the bounding boxes dilation (#128) (82372a1)

Feature

• Add predict_batch to layout predictor (#125) (93ce0ba)

Fix

• Use device_map for transformer models (#124) (58656c3)
• Improve calculation of num_cols and num_rows (#126) (a45d146)

Fix

• Table cell alignment regression (#122) (389161c)

Feature

• Refactor the LayoutPredictor to support all layout models (#121) (505fbf4)

Feature

• Add enumerated field inference to ListItemMarkerProcessor (#119) (a7fa2b8)

Fix

• Secure torch model inits with global locks (#120) (bfef09c)

Feature

• Add initial rule-based model to identify ListItem markers (#113) (e063b97)

Feature

• Performance optimizations for reading order and table model (#115) (0758ad1)

Fix

• Remove deps constraints (#111) (b2c091f)

Fix

• Python3.13 dependencies compatibility (#91) (3adaf74)

Fix

• Table model - optimizing align_table_cells_to_pdf in matching_post_cessor (#93) (6b7b036)

Fix

• Remove regex warning in reading_order model (#84) (e22095b)

Feature

• Add readingorder model (#44) (23c1696)

Fix

• Add CodeItem to caption linking algorithm in reading-order-model (#81) (c3b53a5)

Fix

• Update Pillow constraints (#80) (4fa1828)

Fix

• Performance issue code formula model (#78) (b78d81f)
• Upgrade to code formula model v1.0.1 (#75) (d6a3549)
• Enable MyPy in pre-commit and refactor the code to fix all errors (#74) (a47673b)

Feature

• New document figure classifier model (#73) (60807a7)

Fix

• Fixed prompt of code formula predictor (#72) (bdcc82f)

Feature

• Code equation model (#71) (fa51a6c)

Fix

• Use old transformers version with old torch version (#70) (b3e072e)

Fix

• Force numpy < 2.0.0 on mac intel (#69) (7f9365f)

Feature

• Add arguments for LayoutPredictor (#66) (fe6a476)

Feature

• New API for models initialization with accelerators parameters. Use HF implementation for LayoutPredictor. Migrate models to safetensors format. (#50) (04295b2)

Breaking

• New API for models initialization with accelerators parameters. Use HF implementation for LayoutPredictor. Migrate models to safetensors format. (#50) (04295b2)

Fix

• Remove print statements (#63) (da13863)

Fix

• Improve numpy compatibility pinning (#57) (de2f241)

Fix

• Python3.9 support (#54) (e2b19d9)

Fix

• Removing dependency from mean_average_precision package (not in use) (#53) (65affef)

Fix

• Remove lxml deps (#51) (7a0cbde)

Fix

• Simplify torch dependencies in the wheels (#45) (bca09f8)

Fix

• LayoutPredictor: Ensure that the predicted bboxes are minmaxed inside the image boundaries (#42) (216cee0)

Fix

• Numpy with python 3.13 support (#39) (4fddc45)

Feature

• Release v2.0.0 with only torch models (#38) (8719555)

Breaking

• release v2.0.0 with only torch models (#38) (8719555)

Warning. This release have been moved to v2.0.0

Feature

• Migration from onnx to pytorch script (#37) (59e2941)

Fix

• Put back in common.py the function read_config(). Extend the unit tests. (#36) (d0bdb22)

Fix

• Remove left-over code which is not needed for prediction (#35) (b6ba0c7)

Fix

• Pinned opencv-python-headless to version "4.6.0.66" (#34) (484340f)

Feature

• Extend the tests and demo to first download the model files from HF. Add the pytest in GitHub workflow ( #30) (79888d0)

Fix

• Safer bbox processing (#27) (d37272e)
• Make col/row re-sorting optional on TF predictor (#19) [skip ci] (7e9758c)

Feature

• LayoutPredictor: Introduce black-listed classes which are filtered out from the response. (#26) (86a6a50)

Fix

• Validation and typechecks in TF post processing and OTSL to HTML conversion function (#18) (d607914)

Fix

• TableFormer raises IndexError: too many indices for array (ad494ca)

Fix

• poetry: Remove unused dependencies from toml. Update lock. (#16) (3792577)

Fix

• Fix torch dependency for Intel Macs (#15) (12153f8)

Fix

• Table cell overlap removal in TF post-processing: (#10) (e8f396d)

Fix

• Align to use opencv-python-headless (#12) (22097fb)