Access is granted automatically. Please tell us who you are and how you intend to use this resource so we can support your work and share licensing options. By requesting access you agree not to redistribute the data and to contact info@nalandadata.ai for commercial licensing.

DrishtiTable: A Benchmark for Table Structure Recognition in Indian Academic Textbooks

🏆 Live Leaderboard: DrishtiTable Leaderboard — see how frontier models rank on this benchmark.

🔬 Try the model: DrishtiTable Live Demo — upload any table image, get HTML back.

📦 Public sample: Nalandadata/DrishtiTable-sample — 10 rows, no login required.

License / commercial access: request via the gated access button above, or contact us. For licensing: info@nalandadata.ai

This is a sample release containing 20 representative tables (10 train / 5 val / 5 test). The full dataset contains 1,421 tables; the 135-table benchmark test set is held out. Contact the authors for full dataset / test-image access.

DrishtiTable is a curated benchmark of table images with high-quality HTML structure annotations from Indian academic textbooks (S. Chand Publications). It evaluates Table Structure Recognition (TSR) — converting a table image into machine-readable HTML structure and content — on domain-specific educational content that commercial models have not seen.

🏆 Benchmark Results (135-table held-out test set)

Ranked by TEDS (Tree-Edit-Distance Similarity, the field-standard TSR metric).

Rank	Model	Method	TEDS	S-TEDS
🥇	DrishtiTable-Qwen2.5-VL-7B	SFT (QLoRA)	83.2%	89.7%
🥈	Claude Sonnet 4.6	Zero-shot	77.3%	89.2%
🥉	Claude Opus 4.8	Zero-shot	75.5%	88.2%
4	GPT-4o	Zero-shot	71.1%	84.3%
5	GPT-5.1	Zero-shot	69.9%	83.3%
6	GPT-4.1	Zero-shot	68.0%	80.8%
7	Gemini 3.1 Pro	Zero-shot	65.7%	73.6%
8	GPT-5 mini	Zero-shot	62.2%	72.7%
9	o4-mini	Zero-shot	61.4%	70.0%
10	Qwen2.5-VL-7B	Zero-shot	58.8%	74.0%

The fine-tuned 7B model beats every zero-shot frontier model, including Claude and GPT-5.1, on this domain-specific benchmark. Full per-model and per-table-type numbers: benchmark/leaderboard.md.

Want your model on the board? See SUBMISSION.md.

📏 The Metric — TEDS

DrishtiTable scores predictions with TEDS, the standard metric for image-based table recognition (introduced by IBM for PubTabNet, ECCV 2020, and used by PubTabNet / FinTabNet / SciTSR / ICDAR 2021). It parses predicted and ground-truth HTML into trees and computes a normalized tree-edit similarity on a 0–100% scale, giving partial credit for partially-correct tables.

Metric	Meaning
TEDS	Structure + cell content similarity (headline / ranking metric)
S-TEDS	Structure only (cell text ignored)
Exact-match	% of tables reconstructed exactly (100%)
Near-perfect	% of tables scored ≥ 95%

The canonical scorer is scripts/score_submission.py — the exact metric behind every leaderboard row.

📖 Full transparency — read or re-implement the benchmark yourself

Document	What's inside
`benchmark/METHODOLOGY.md`	Complete scoring method: TEDS step-by-step, terminology glossary, worked examples (real data), the four metrics, per-category scoring, and an end-to-end reproduce-it-yourself guide.
`benchmark/FAILURE_MODES.md`	Failure taxonomy: score-band distributions, per-model failure signatures, failure-by-table-type, all from real predictions.
`benchmark/BENCHMARK.md`	The formal v1.0 specification (task, protocol, rules, versioning).
`benchmark/PROMPT.txt`	The exact standard prompt.
`scripts/reproduce_score.py`	Run it to see TEDS/S-TEDS computed on tiny tables before scoring a real model.

Everything an implementer needs — scoring terminology, methodology, failure modes, and reproduction steps — is open in this repo. pip install apted beautifulsoup4 and you can recompute every number on the leaderboard.

🔒 Held-out test set (benchmark integrity)

The 135-table test split is frozen as v1.0 and content-hashed in benchmark/test_manifest.json (SHA-256), so all leaderboard numbers are permanently comparable. The test ground-truth HTML is held out to prevent training-on-test; test images are available on request. Submissions are scored manually by the maintainers against the held-out labels.

What's public vs. private

Asset	Public on the Hub	Private / on request
Tables (image + HTML label + metadata)	20 (~1.4%)	1,401 (~98.6%)
Train / Val / Test breakdown	10 / 5 / 5 sample	1,131 / 140 / 130
135 benchmark test labels	0 (held out)	all 135 (used for scoring)

The public test.csv lists only the 5-table sample. The full 135-table test set is scored by the maintainers against the private labels. Request the test images to self-evaluate: tech@nalandadata.ai.

Sample Dataset Statistics

Split	Samples (this release)	Samples (full dataset)
Train	10	1,141
Validation	5	145
Test (benchmark)	5	135
Total	20	1,421

Full Dataset Statistics

The complete DrishtiTable dataset spans 9 Indian academic textbooks across 6 subjects:

Subject	Tables
Financial Accounting	780
Business Statistics	261
Quantitative Techniques	183
Operation Research	108
Steam Tables (Engineering)	52
Ethics	37

Table-type taxonomy: Statistical · Financial · Lookup · Comparison.

Complexity Features (Full Dataset)

Feature	Count	Percentage
Merged cells (colspan/rowspan)	200	14.1%
Multi-level hierarchy	219	15.4%
Empty cells	340	23.9%
Bold text	890	62.6%

Data Format

Each sample consists of:

Image: Cropped table region as PNG (images/)
Annotation: HTML structure with semantic tags (annotations/)
Metadata: Table properties as JSON (metadata/)

HTML Annotation Example

<table>
  <thead>
    <tr>
      <th>Year</th>
      <th colspan="2">Revenue</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>2024</td>
      <td>1,234</td>
      <td>5,678</td>
    </tr>
  </tbody>
</table>

File Structure

DrishtiTable/
  images/                       # PNG table images (20-table sample)
  annotations/                  # HTML ground-truth files
  metadata/                     # JSON metadata files
  train.csv / val.csv / test.csv
  benchmark/
    BENCHMARK.md                # full benchmark specification (v1.0)
    PROMPT.txt                  # standard TSR prompt
    test_manifest.json          # frozen 135-table test set + SHA-256
    leaderboard.md              # static leaderboard (human-readable)
    leaderboard.jsonl           # leaderboard data (machine-readable, source of truth)
  scripts/
    score_submission.py         # canonical TEDS scorer
    build_leaderboard.py        # regenerates the board from results
    freeze_test_manifest.py     # freezes the test split
  SUBMISSION.md                 # how to get on the leaderboard

How to evaluate / submit

pip install apted beautifulsoup4
# Score a predictions file ({table_id: html}) with the official metric:
python scripts/score_submission.py \
    --pred predictions.json \
    --annotations annotations/ --test-csv splits/test.csv \
    --model "Your Model" --org "Your Org" --method "Zero-shot"

See SUBMISSION.md for the full process.

Fine-Tuned Model, Demo, and Leaderboard

Resource	Link
🏆 Leaderboard	DrishtiTable Leaderboard
🔬 Live Demo	DrishtiTable Space
Fine-tuned Model	Nalandadata/DrishtiTable-Qwen2.5-VL-7B
Base Model	Qwen/Qwen2.5-VL-7B-Instruct

Citation

@article{drishtitable2026,
  title={Domain-Specific Fine-Tuning for Table Structure Recognition: A 7B Open Model Outperforms GPT-4o with 1,141 Training Samples},
  author={Nalanda Data},
  year={2026}
}

TEDS metric:

@inproceedings{zhong2020image,
  title={Image-based table recognition: data, model, and evaluation},
  author={Zhong, Xu and ShafieiBavani, Elaheh and Jimeno Yepes, Antonio},
  booktitle={ECCV},
  year={2020}
}

Commercial Use & Licensing

This dataset is released under Apache 2.0, which permits commercial use under the terms of that license.

The public release on Hugging Face is a sample of 20 tables. The full DrishtiTable corpus contains 1,421 expert-annotated tables across 9 Indian academic textbooks, spanning Financial Accounting, Business Statistics, Quantitative Techniques, Operations Research, Steam Tables, and Ethics.

For commercial users who need:

Full dataset access — 1,421 tables across 6 subjects
Benchmark test images — the 135-table held-out evaluation set
Custom slices — specific subjects, table complexity profiles, multi-page layouts
Custom data work — additional textbook collection, annotation, evaluation pipelines
Production deployment support for fine-tuned TSR models