Dataset Card — indic-reg-bench

Status: under construction. The gold set does not exist yet. This card describes what is built, what is not, and the decisions taken so far. It will be wrong in places until labelling is finished; it is published early so the construction method is auditable rather than reconstructed afterwards.

What this is

A benchmark for Indian regulatory document understanding, built on SEBI (Securities and Exchange Board of India) enforcement orders. Five tasks: structured extraction, charging-section prediction, numeric reasoning, abstention, and penalty attribution. See docs/phase1-task-design.md.

What is actually in it right now

Component	State
Listing metadata	11,957 orders, Nov 2004 – Jul 2026, scraped from sebi.gov.in
Fetched document text	25 orders (pilot), 1,476,418 chars
Gold labels	none — 0 examples
Silver labels	none
Splits	temporal, defined (train <2023 / test ≥2023); not yet populated
Harness	working, pip-installable, 15 tests passing
Baselines	regex floor implemented; no scores, because scoring needs gold labels

Provenance

Source: https://www.sebi.gov.in/enforcement/orders/..., the public enforcement-orders section. Listing metadata is collected via the site's own AJAX paging endpoint; each order's PDF is resolved from the viewer iframe on its page and text-extracted with pdfplumber.

Documents were retrieved between 2026-08-05 and the date in each row's fetched_at. SEBI may revise or withdraw orders; a SHA-256 per document is planned so drift is detectable.

Redistribution and licence

The source order text is not redistributed. This repository ships document IDs, source URLs, and a fetch script. Anyone can rebuild the corpus from SEBI's servers with one command.

This is a deliberately conservative position. SEBI orders are public documents and Indian copyright law treats government works distinctly, but rather than assert a redistribution right I am not qualified to assert, the benchmark keeps SEBI as the authoritative source. The harness, scorers and labels are Apache-2.0; the order text is SEBI's and is not covered by that licence.

Cost of this choice: reproducibility depends on SEBI keeping URLs live, and a fetch takes hours. That is a real downside and it is accepted knowingly.

Personal data

SEBI enforcement orders name private individuals and contain, variously, addresses, PAN references, trading account details, and family relationships. The pilot sample includes three orders concerning deceased persons (Late Ms. Anju Rani, Late Padma Singhwani, Late Sudha V Thakkar), where proceedings continue against legal heirs who are themselves named.

Consequences, stated plainly:

• Noticee names are the labels for T1. A version with names removed would not be a usable extraction benchmark.
• The current intention is to publish names, since they are already public in the source documents. This decision is not yet final and is flagged for the maintainer.
• No attempt is made to enrich, cross-reference, or link these individuals to any other dataset, and doing so is out of scope.
• A takedown contact will be published before any HuggingFace release.

This dataset must not be used to profile, score, or make decisions about the named individuals. It exists to measure document-understanding systems.

Known biases and limitations

• One regulator, one document type. SEBI adjudication orders only. Not RBI, not IRDAI, not tribunal judgments. Results do not generalise to Indian regulatory text broadly.
• Extraction-heavy. Four of five tasks reward locating and normalising text. Legal reasoning is barely tested.
• Single annotator. Agreement will be reported as self-agreement across two passes a week apart. That is an upper bound on reliability: the same reader reproduces their own systematic misreadings and scores them as agreement.
• English only. SEBI publishes in English; Indian-language regulatory text is untested.
• Class imbalance. Charging sections are dominated by 15HB; the pilot shows 27 mentions of 15HB against 16 of 15HA and 8–9 each of 15EA/15EB. Macro-F1 and a published majority-class baseline are used for this reason.
• Multi-noticee orders are rare — 3 of 25 in the pilot. They are the interesting case, so the sample deliberately oversamples them; the resulting distribution is not representative of SEBI's output and must not be read as such.
• Multi-hop is absent from v1. Only 9 of 133 entities in the pilot recur across documents, and all 9 come from a single matter family. The task was cut rather than shipped degenerate.
• OCR/extraction noise is real. Penalty tables are column-scrambled by text extraction; the annotation guidelines require labelling such orders from the source PDF.

What this benchmark does not measure

Legal correctness. Whether SEBI's reasoning is sound, whether a penalty is proportionate, or whether an order would survive appeal. It measures whether a system can read what an order says.

Construction method

1. Scrape listing metadata for all 11,957 orders (scripts/scrape_listing.py).
2. Select a stratified sample by year, adjudication orders only, seed 42 (scripts/build_splits.py).
3. Fetch and text-extract (scripts/fetch_orders.py).
4. Hand-label the gold set with scripts/label.py, logging every decision to labels/decisions.jsonl.
5. Re-label 50 orders after ≥7 days; report self-agreement (scripts/agreement.py).
6. Cut temporal splits; run baselines; publish per-task results and a failure taxonomy.

Steps 1–3 are built. Step 4 onward is not started.

Citation

@misc{indicregbench2026,
  title  = {indic-reg-bench: A Benchmark for Indian Regulatory Document Understanding},
  author = {Gaur, Siddharth},
  year   = {2026},
  url    = {https://github.com/siddharthgaur1/indic-reg-bench}
}

---

Files in this release

• listing.csv — 11,957 SEBI enforcement orders (url, date, year, title, doc_type). The complete index, Nov 2004 – Jul 2026.
• benchmark_sample.csv — the 2,000 adjudication orders stratified by year (seed 42) that the benchmark is being built from.

There is no order text and there are no labels in this release. Fetch the documents yourself:

git clone https://github.com/siddharthgaur1/indic-reg-bench
cd indic-reg-bench && pip install -e .
python scripts/scrape_listing.py
python scripts/fetch_orders.py --fetch-set