tablereader-lightonocr-v0

LightOnOCR-3-4B fine-tuned to transcribe a crop of a printed table into a fixed HTML format: a # Title line when a title is printed, a table with thead for printed column headings, data-level for the printed indentation of row labels, <sup> for footnote markers and tfoot for footnote lines (transcription format, contract revision 1).

It is the model behind tablereader, which detects the tables in a PDF, crops them over their titles and footnotes, runs this model, and converts the transcription into table JSON. release.json records the settings the model was trained with: crop preparation, renderer, detector, image size and generation settings. Use it through tablereader so crops are prepared the same way:

uv run tablereader read report.pdf --pages 13-20

Prompt

The image alone, with no system prompt or text, scaled so its longest edge is at most 2048 pixels; thinking disabled, temperature 0, up to 16,000 new tokens.

Training

LoRA (rank 16, alpha 32) on every language-model projection, the vision encoder frozen; 3 epochs at learning rate 1e-4 on 995 table crops from 107 publication series (UK public-sector reports, US federal and historical statistics, German budget annexes and others), digital and scanned. The targets are GPT-6.1 Terra transcriptions that parse into tables without blocking issues; they were not reviewed by a person. The training crops include some that the scannedtables dataset holds back for rights or privacy reasons. The adapter is merged into the bf16 weights.

Evaluation

1,006 validation crops from series not seen in training, scored as the share of printed values whose row path, column path and text match the reference transcription exactly (metric records-1):

All	Digital PDFs	Scans
0.797	0.835	0.511

72% of crops are read without a difference. The model does not yet answer NOT_A_TABLE for crops without a table, and scans read much worse than digital PDFs. A self-training round on 7,698 of its own transcriptions did not improve these scores.

License

Apache 2.0, as the base model.