# SynthTabNet

10,152 training samples. One source, one `train.parquet`.

Each row embeds original image `bytes`, `width`, `height`, authoritative `otsl`, derived `html`, logical `rows`, `cols`, `sample_id`, `document_id`, image/OTSL SHA256, `structure_only`, and `model_budget_eligible`.

No resizing, re-encoding or model prediction is used. Dimensions are read from image bytes; HTML is derived from OTSL. Original annotation review does not imply every sample was manually checked.

## PaddleOCR-VL-1.6 training-pool evaluation

The single root-level [paddleocrvl1.6_eval.json](../paddleocrvl1.6_eval.json) contains one JSON entry per Parquet sample, in source/Parquet order. **10,152/10,152** have TEDS-S scores from the original base model. Match using `sample_id` and `gt_sha256` (`otsl_sha256` in Parquet). Unscored metrics are `null`; textless GT has `teds=null`. Scores use the project canonical scorer and existing selected screening subset, not an official held-out benchmark.