This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence (CC BY-NC-SA 4.0): research, academic, personal and other non-commercial use is permitted with attribution; adaptations must be shared under the same licence; commercial use is not permitted. The original public source data remains under the Government Open Data License - India (GODL-India). Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

IndiaTables-Bench

Licence: CC BY-NC-SA 4.0 — research and non-commercial use only; derivatives must be shared under the same licence. Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai

Access is approved automatically when the email you enter matches your Hugging Face account email and belongs to the organisation you name; other requests are reviewed manually.

Five tasks over Indian government tables from IndiaTables (data.gov.in and NITI Aayog NDAP, snapshot May 2026): metadata extraction from the grid alone, column semantic types, schema matching, header-tree reconstruction and table retrieval with sibling distractors. Splits are by template family (no column layout spans two splits) with whole ministries held out for test (Department of Water Resources, River Development & Ganga Rejuvenation, Karnataka, Tamil Nadu). At most 3 tables per template. evaluate.py scores predictions and reproduces the reference numbers below.

Headline result: on task 5 (table retrieval) IndiaTables-Embed-0.6B reaches nDCG@10 0.8388, against 0.6825 for pplx-embed-v1-0.6b, 0.6208 for BAAI/bge-m3 and 0.4900 for BM25.

Task	train	dev	test
task1_metadata	42,370	2,602	4,752
task2_coltypes	378,791	23,184	40,966
task3_schema_match	1,557	84	183
task4_header_tree	13,807	840	1,521
task5_retrieval	214,541	13,292	23,961

Metadata extraction: column names + 5 rows, no title → what was measured (title), geographic level, years, ministry.
Column types: one column (header + up to 8 values) → one of 32 types (serial_no, state, district, ..., amount_lakh, percent, other_numeric). Labels come from rules on the header and typed values.
Schema matching: two tables of one template family with different layouts → column correspondences.
Header tree: flattened A - B - C headers → the nested header tree.
Table retrieval: a question (English, Hindi or Hinglish) → the table that answers it, out of all 266,707 IndiaTables cards. Questions were generated from table rows and kept only when the answer value occurs in the table; sibling tables (same report, other district or month) are the main distractors.

Quickstart

Input: a task's JSONL split. Output: metric dicts printed by the scorer.

import json, sys
from huggingface_hub import snapshot_download

root = snapshot_download("judicialmind/IndiaTables-Bench", repo_type="dataset",
                         allow_patterns=["evaluate.py", "bench/*"])
sys.path.insert(0, root)
from evaluate import read_jsonl, score, prompt

items = read_jsonl(f"{root}/bench/task2_coltypes/test.jsonl")
print(len(items), items[0]["input"]["header"], items[0]["input"]["values"][:2], "->", items[0]["target"])
print(prompt("task2_coltypes", items[0])[:80])

preds = {it["id"]: "count" for it in items}          # replace with your model's labels
print(score("task2_coltypes", items, preds))

Expected output: 40,966 test items; the first has header State, values ['Andaman & Nicobar', 'Andaman & Nicobar'] and target state. The first 80 characters of its prompt:

Column header: State
Values: Andaman & Nicobar, Andaman & Nicobar, Andaman & Nic

Constant-"count" baseline on task2_coltypes (test)

Accuracy	Macro F1	Items
0.0865	0.0051	40,966

Command line (run inside the downloaded folder):

python evaluate.py baselines                                   # re-score the shipped reference predictions
python evaluate.py prompts --task task4_header_tree --out prompts.jsonl     # zero-shot prompts for your LLM
python evaluate.py score --task task4_header_tree --pred my_preds.jsonl     # {"id", "pred"} per line

# task 5 needs the IndiaTables corpus files
huggingface-cli download judicialmind/IndiaTables cards.parquet catalog.parquet --repo-type dataset --local-dir corpus
python evaluate.py retrieval --cards corpus/cards.parquet --catalog corpus/catalog.parquet --bm25
python evaluate.py retrieval --cards corpus/cards.parquet --catalog corpus/catalog.parquet \
    --model judicialmind/IndiaTables-Embed-0.6B --write-run run.jsonl  # any sentence-transformers model
python evaluate.py retrieval --cards corpus/cards.parquet --catalog corpus/catalog.parquet --run my_run.jsonl

Prediction formats (pred): task 1 {"title", "level", "years", "ministry"}; task 2 a type label; task 3 [[a_column, b_column], ...]; task 4 a nested dict with {} leaves. parse_reply(task, text) in evaluate.py turns a raw LLM reply into this format. A retrieval run is JSONL {"id": <query id>, "ranking": [uuid, ...]}, best first (top 100 is enough).

To submit a result, open a discussion on this repository with your scores from evaluate.py and the prediction or run file.

Files

File	What
`bench/<task>/{train,dev,test}.jsonl`	Tasks 1-4
`retrieval/{train,dev,test}.jsonl`	Task 5 questions
`bench/preds/qwen3-8b/<task>.test.jsonl`	Reference zero-shot predictions (tasks 2-4)
`bench/splits_meta.json`	Held-out ministries and item counts
`bench/_split_map.json`	Split of every IndiaTables table (`uuid -> train/dev/test`); keep test tables out of training
`evaluate.py`	Scoring, prompts, BM25 and dense retrieval runs

Schema

Tasks 1-4 (one item per line): id, uuid (IndiaTables table), split, ministry, template_id, template_family, source, input, target; task 2 also column and label_source; task 3 also other_uuid.

Task	`input`	`target`
task1_metadata	grid text: header line + 5 rows, `	`-separated
task2_coltypes	`{"header", "values"}`	type label
task3_schema_match	`{"a": [columns], "b": [columns]}`	`[[a_column, b_column], ...]`
task4_header_tree	list of flattened headers	nested dict, `{}` leaves

Task 5 (retrieval/*.jsonl): id, query, lang (en, hi, hinglish), positives (uuids), hard_negatives (top-ranked non-answer tables from the same split), sibling_negatives (siblings that the question rules out), template_id, ministry, split, generator, checked, check, answer_value, answer_column (the cell that answers the question).

Results

Test split. Task 5 corpus: all IndiaTables cards (266,707 tables).

Task 5, table retrieval:

model	nDCG@10	R@1	R@10	MRR@10
bm25	0.4900	0.3514	0.6335	0.4443
BAAI/bge-m3	0.6208	0.4762	0.7712	0.5729
Qwen/Qwen3-Embedding-0.6B	0.5808	0.4378	0.7336	0.5324
perplexity-ai/pplx-embed-v1-0.6b	0.6825	0.5415	0.8238	0.6373
IndiaTables-Embed-0.6B	0.8388	0.7225	0.9455	0.8041

Tasks 2-4, Qwen3-8B zero-shot with the prompts in evaluate.py (predictions in bench/preds/qwen3-8b/):

Task	metric	Qwen3-8B
task2_coltypes	accuracy	0.3968
task2_coltypes	macro F1	0.4497
task3_schema_match	micro F1	0.5375
task4_header_tree	edge F1	0.4059
task4_header_tree	exact match	0.1374

Coverage

Tables from all ministries in IndiaTables (data.gov.in and NDAP, snapshot May 2026), national to village level; task 5 questions in English, Hindi (Devanagari) and Hinglish.

Source data

Built from IndiaTables: tables and catalog metadata published on the Open Government Data Platform India (data.gov.in) and NITI Aayog's NDAP, snapshot May 2026. Every item carries the uuid of its source table; the per-table GODL-India attribution is in the IndiaTables catalog.parquet attribution column.

Original data: published by the ministries, departments and State governments named per table in IndiaTables on the Open Government Data Platform India (data.gov.in) and, where marked, NITI Aayog's National Data and Analytics Platform (NDAP), under the Government Open Data License - India (GODL-India). The original public data remains under GODL-India; the CC BY-NC-SA 4.0 licence of this repository covers JudicialMind's compilation, derived fields, annotations, benchmarks and models.

Citation

@misc{judicialmind2026indiatablesbench,
  title={IndiaTables-Bench: five tasks over Indian government data tables},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/IndiaTables-Bench}}
}

License

CC BY-NC-SA 4.0 (see LICENSE): research, academic, personal and other non-commercial use, with attribution to JudicialMind and to the original data publishers; adaptations must be shared under the same licence. Commercial use requires a separate licence.

Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

Downloads last month: 19

Total file size:

476 MB

Collection including judicialmind/IndiaTables-Bench

India Open Government Data — Bharat Tables Benchmarks

Collection

Research and non-commercial use (CC BY-NC-SA 4.0). Commercial: surya@judicialmind.ai / hello@judicialmind.ai • 3 items • Updated 1 day ago