CheneyWong/table_recognition_train_canonical

Table recognition training — one source, one Parquet

The frozen training pool contains 1,454,538 samples across 9 sources. 9/9 source folders are uploaded and remotely verified.

Each source has its own folder with exactly one train.parquet, audit.json and README.md. Original encoded image bytes are embedded; there is no dependency on local ZIP files.

Source / folder	Samples	Upload status
FinTabNet_OTSL	81,032	verified
MMSci	33,336	verified
PubTabNet	468,698	verified
PubTables-1M	728,203	verified
PubTables-v2	104,219	verified
SciTSR	8,228	verified
SynthTabNet	10,152	verified
TABLET	16,822	verified
synth_2d	3,848	verified

Schema

Field	Meaning
bytes	Original encoded image, unchanged
width / height	Actual image dimensions read from image bytes
otsl	Authoritative reviewed-pool OTSL, unchanged
html	Derived from OTSL; entities decoded and cell text HTML-escaped; not a model prediction
rows / cols	Logical table grid dimensions reconstructed from OTSL
source / sample_id / document_id	Source and stable sample/document identity
image_sha256 / otsl_sha256	Integrity checks against the audited training master
structure_only	No cell text; structure-only supervision
model_budget_eligible	Whether the sample fits the current model token budget; all valid pool rows are retained
html_origin	HTML conversion provenance

TabRecSet zh/en, cxrender, cxrender_zh, WTW and real-scene training data are excluded according to the active pool policy. Upstream split and held-out/evaluation exclusions are preserved. This is the unified pool from which SFT/RL/OPD selections are made, rather than one stage's selected subset. Annotation auditing does not mean every image was manually reviewed.

Loading

from datasets import load_dataset
from PIL import Image
import io
ds = load_dataset("CheneyWong/table_recognition_train_canonical", "PubTables-1M", split="train")
sample = ds[0]
image = Image.open(io.BytesIO(sample["bytes"]))
# Default config combines all uploaded sources.

Each source audit includes the Parquet SHA256, sample count and originating master SHA256. All image and OTSL hashes are checked after writing Parquet, and every uploaded Parquet is compared with its remote LFS SHA256. No resizing, re-encoding, new training or model inference is performed during this export.

Original datasets retain their provenance and original terms; this repository reorganizes the supplied audited training pool.

PaddleOCR-VL-1.6 training-pool evaluation

paddleocrvl1.6_eval.json is the single aggregate JSON containing provenance, metric definitions, per-source coverage and a samples array covering every training Parquet row.

175,681/1,454,538 training samples currently have TEDS-S scores; 1,278,857 have no cached inference. These are original PaddleOCR-VL-1.6 base-model greedy screening results, never SFT/RL/OPD checkpoint results. The scored subset was selected for screening and is not an unbiased full-pool benchmark.

JSON fields include source, sample_id, gt_sha256, teds, teds_s, status, scoring_status and finish_reason where applicable. Join by sample_id within each source and verify gt_sha256 == otsl_sha256. not_evaluated means both scores are null, not zero. For GT without cell text, teds=null and only TEDS-S applies. Numeric metrics are in [0,1]. Scoring uses the project canonical TEDS implementation with reviewed GT and text normalization; it is not the separate official fullpage benchmark. Training Parquets and original images are unchanged.

import json
from huggingface_hub import hf_hub_download
p = hf_hub_download("CheneyWong/table_recognition_train_canonical", "paddleocrvl1.6_eval.json", repo_type="dataset")
evaluation = json.load(open(p))
by_id = {r["sample_id"]: r for r in evaluation["samples"]}