ocr-to-typst
Star
Fork
Watch

Commits on canon
Update README
committed
augment_vis: invert handwriting splits
committed
Add dark-mode inversion augmentation for handwriting splits
committed
Fix dark-mode augmentation and simplify augment-vis
committed
Track regenerated typeset splits with dark mode
committed
Overhaul augmentation pipeline and add augment-vis tool
committed
Add typeset_prose splits and augmentation improvements
committed
Add deepseek handwriting demo web app
committed
Add DeepSeek-OCR-2 model code with eval_mode inference fix
committed
Move deepseek-ocr 2 finetune code to subdirectory to use earlier version of transformers
committed
WIP: new deepseek-ocr 2 impl
committed
Remove previous deepseek-ocr 2 finetune attempt
committed
Add image-stats command
committed
Regen typeset splits at 250 PPI; rename structured->uniform; strengthen prompt
committed
Add licenses; flatten fonts dir; update font path references
committed
Rename hw_* splits to typeset_structured/mixed; drop old typeset_* splits; pin unsloth
committed
Update README: reflect hw_* splits, new body grammar weights, updated data mix
committed
Add handwriting-font data pipeline; expand body grammar; drop typeset_* splits
committed
Update eval (per-split sampling, verbose mode); align save_steps with eval; drop DeepSeek from README
committed
Add crohme_gen_syntactic split; fix val stratification; reduce batch size; fix checkpoint resume; drop transformers pin
committed
Overhaul Gemma training; fix eval stratification; fix mathwriting label format
committed
Fix image injection bug; add DeepSeek eval/mining scripts
committed
Add TensorBoard logging; backfill script for stdout logs
committed
Fix training loss logging; report_to=[] instead of 'none'
committed
Increase batch size to 2, halve grad_accum; update seed to 29979
committed
Cap typeset_val at 250; document training mix and splits in README
committed
Rename project to typst-ocr
committed
Switch to Python 3.12; 1 epoch; rebalance training caps
committed
Add README with training split documentation; rebalance train data; fix val
committed
Remove debug prints from modeling_deepseekocr2.py
committed
Fix DeepSeek-OCR-2 for training; load from local patched module
committed
Add DeepSeek-OCR-2 modeling files from HuggingFace hub (verbatim)
committed
Add train_deepseek: QLoRA fine-tuning script for DeepSeek-OCR-2
committed
Add DeepSeek-OCR-2 data collator
committed
Add data/typeset_mixed_*
committed
Wrap math-only split labels with display math delimiters at load time
committed
Add underline to text groups; reduce styling rates to 25%
committed
Add polynomials, schematic matrix, multi-paragraph, and mixed list+para bodies
committed
Fix list item overlap from tall math expressions
committed
Add row vectors and non-square matrices to grammar
committed
Expand grammar: cal(), logic/sequent, vec, overbrace, tables, bra-ket, intervals
committed
WIP: Add generate_mixed dataset for more structure
committed
Add replace-dif script and apply d→dif label rewrites
committed
Add boolean search and result counts to label review
committed
label app: Add search
committed
Add replace_frac
committed
Replace deprecated 'diff' keyword with 'partial' in labels
committed
Replace frac(A,B) with (A)/(B) in all split labels
committed
Update 4 labels from crohme_real_train (gt/lt => >/<)
committed
Canonicalize ^prime -> ' in labels (dataset-v2 prep)
committed
search-labels: add --raw flag for literal substring patterns
committed
Add search-labels CLI for inspecting and bulk-replacing label patterns
committed
Add label review web app and apply-edits CLI
committed
Add vim swap files to .gitignore
committed
Move probe_deepseek to src package, drop stale scripts/ copy
committed
Consolidate dataset under data/, init DVC, track all 14 splits as v1
committed
Add DeepSeek-OCR-2 standalone scripts and project deps
committed
Bump max_new_tokens to 512 and add raw output debug logging
committed
Add FastAPI web app and refine probe/eval utilities
committed
Import eff-mer CNN+Transformer model for comparison baseline
committed