indic-dataset-license-matrix / sample-audit-report.md
hardik90's picture
Upload sample-audit-report.md with huggingface_hub
65462f3 verified
|
Raw
History Blame Contribute Delete
3.45 kB

Sample Audit Report — Indic LLM Training Data Compliance

Prepared by: Independent Hub Auditor (license-audit service) Scope: 3 flagship Indic corpora commonly used for LLM training Scan date: 2026-08-01 | Confidentiality: findings are public metadata; analysis and recommendations are proprietary


1. Executive summary

Dataset Downloads Declared tag Risk level Bottom line
ai4bharat/IndicCorpV2 2,888 none HIGH Not verifiable; not usable in commercial training without upstream clearance
cfilt/iitb-english-hindi 1,003 none HIGH Benchmark use OK for research; commercial redistribution unclear
ai4bharat/IndicParaphrase 4,611 CC-BY-NC-4.0 HIGH (restricted) Research only; commercial training prohibited

Observation: the most-downloaded Indic corpora carry the weakest license documentation. Under the EU AI Act (GPAI training-data documentation duties, Article 53) and typical enterprise open-source policies, these datasets are effectively excluded from commercial pipelines until provenance is verified.

2. Detailed findings

2.1 ai4bharat/IndicCorpV2 — no license tag

  • Hub tag: absent. Card: absent. No LICENSE file in repo.
  • Public context: ai4bharat's corpus website historically states non-commercial research intent for some corpora. The Hub presence gives no signal whatsoever — a downstream user cannot distinguish "permissively licensed" from "all rights reserved."
  • Mitigations available: (a) contact maintainers for written terms; (b) check the original corpus site + published paper; (c) substitute with a license-verified corpus (see recommendations).

2.2 cfilt/iitb-english-hindi — no license tag

  • Same gap. This corpus is used in WAT translation shared tasks; academic users are safe, commercial redistribution of derivative fine-tunes is undocumented.
  • Even benchmark evals used in a commercial product review should log provenance.

2.3 ai4bharat/IndicParaphrase — CC-BY-NC-4.0

  • Tag is correct and machine-readable — good.
  • NC restriction is disqualifying for commercial fine-tuning or RAG corpora. Common failure: teams fine-tune on it because "it's from ai4bharat, it's open."
  • Compliance check: confirm no synthetic augmentation of this data was used in any shipped model, or you are in breach of the license grant.

3. Recommended remediation

  1. Immediate: add a watchlist flag for all no-license corpora in your training-data inventory (the companion License Matrix CSV lists all 3,185).
  2. 30 days: contact maintainers with a written-use questionnaire for the corpora you actually use; obtain an email grant or a re-release under an OSI license.
  3. Ongoing: log provenance (repo id, commit sha, license tag, retrieved date) per dataset — this is also your EU AI Act Article 53 documentation trail.

4. What a full audit includes

  • Inventory of your actual training/eval datasets with per-repo license verification.
  • Upstream-source chain analysis (translations, scrapes, derived corpora).
  • Written-use recommendations mapped to your jurisdiction (EU AI Act, US, India DPDP).
  • A provenance log template you can hand to counsel.

This sample covers 3 of 4,893 datasets. The full matrix (all rows, CSV) is public on the Hub; a production audit applies the same rigor to your actual inventory.