hardik90/hf-inference-pricing-matrix

Indic Dataset License Matrix

A compliance snapshot of 4,893 Indic-language datasets on the Hugging Face Hub.

Every row is a public dataset discovered via the Hub API (search on: hindi, tamil, bengali, telugu, marathi, malayalam, kannada, gujarati, punjabi, urdu, odia, assamese, nepali, hinglish, code-mixed, indic, indian). This matrix records the declared license tag, a risk bucket, and one-line guidance for commercial use.

⚠️ This dataset describes declared license tags, not verified legal status. A missing tag does not mean "unlicensed by its author" — it means you cannot verify usage rights. Treat this as a triage tool, not legal advice.

Key findings (scan date: 2026-08-01)

Metric	Value
Datasets scanned	4,893
No license tag declared	3,185 (65.1%)
OSI-compatible open license	1,415 (28.9%)
Non-commercial (CC-BY-NC*)	154 (3.1%)
Ambiguous / wrong tag (other, cc, unknown)	139 (2.8%)
Share of all downloads hitting no-license repos	46.1%
Share of all downloads hitting non-commercial repos	7.7%

Notable findings

• ai4bharat/IndicCorpV2 (2,888 downloads) — no license tag on the Hub, despite ai4bharat being the flagship Indic-AI lab. Under EU AI Act GPAI obligations and most enterprise procurement policies, an undocumented corpus is unusable without an audit.
• cfilt/iitb-english-hindi (1,003 downloads, 68 likes) — the classic MT benchmark, no license tag; the original CFILT page carries its own terms, but the Hub tag gives downstream users zero signal.
• Non-commercial traps: ai4bharat/IndicParaphrase (4,611 dl), IndicQuestionGeneration (3,728 dl), IndicHeadlineGeneration (3,475 dl) are all CC-BY-NC-4.0 — safe for research, disqualifying for commercial training.
• Translations (e.g., SmallScale/Simple-Stories-Hindi, Meyank/Tiny_Stories_Hindi) declare licenses inherited from the source corpus. License inheritance is not automatic in law — a translation carries the translator's rights plus upstream terms.

Data dictionary

Column	Meaning
dataset_id	Full repo id on the Hub
author	Uploading user/org
license_tag	Declared license: tag, or (none)
risk_bucket	no-license / non-commercial / unclear / osi-compatible
risk_level	HIGH / MEDIUM / LOW triage
guidance	One-line commercial-use guidance
gated	Whether the repo requires approval
downloads / likes	Usage signals (at scan time)
created / last_modified	ISO dates
description	First 300 chars of the card

Methodology

1. Hub API GET /api/datasets?search=...&full=true for 17 Indic keywords; deduplicated by repo id. 2. License tag extracted from the license: tag list. 3. Bucketing rules: (none) → no-license; cc-by-nc* → non-commercial; unknown/other/cc/gpl/llama2/gemma/ afl-3.0/openrail → unclear; everything else → osi-compatible. 4. Downloads/likes are point-in-time values.

How to use this data

• Dataset owners: find your repo, fix your tag, add a provenance note in your card. Missing tags suppress adoption by procurement teams.
• Enterprises / startups: filter risk_level != LOW to build your watchlist before training or fine-tuning on any Indic corpus.
• Researchers: cite the matrix when describing the Indic-data licensing landscape.

FAQ

Can I use Indic datasets commercially if no license is shown?

No — treat it as all-rights-reserved. A missing tag means you cannot verify usage rights. Under EU AI Act GPAI documentation duties and standard procurement policies, an undocumented corpus is excluded from commercial pipelines until the maintainer confirms terms in writing.

Which Indic datasets are safe for commercial training?

Filter the matrix for risk_bucket == osi-compatible (1,415 repos), then verify the license_tag against the actual LICENSE file — e.g. apache-2.0, mit, cc0-1.0, cc-by-4.0. Watch for translation/derivation chains: a translated dataset is not automatically covered by the source corpus's license.

What does CC-BY-NC mean for fine-tuning?

Non-commercial use only. Fine-tuning a commercial model on CC-BY-NC data breaches the license grant, even if your model is open-sourced afterwards. 154 Indic datasets (7.7% of all downloads) carry NC tags — the top trap is ai4bharat/IndicParaphrase (4,611 downloads).

How do I verify a dataset's license on the Hub?

1. Check the license: tag under the repo header. 2. Look for a LICENSE file in the repo tree. 3. Read the dataset card and the original corpus/paper. 4. If the card claims a license inherited from another dataset, confirm the upstream terms yourself. 5. Record repo id + commit sha + date — that log is your compliance trail.

Why is ai4bharat/IndicCorpV2 flagged as high risk?

It declares no license tag on the Hub despite being one of the most-downloaded Indic corpora (2,888 downloads). The tag absence alone triggers enterprise exclusion rules; maintainers were asked to add documentation (see discussion on that repo).

Audit service

This matrix is a triage layer. If your team trains, fine-tunes, or ships products on any Indic corpus, a written license & provenance audit covers what a tag can't:

• per-repo verification (tag vs. actual LICENSE file vs. upstream source terms)
• translation / derivation chain analysis (e.g. "this Hindi set is a translation of X")
• non-commercial trap detection across your full training inventory
• a provenance log you can hand to counsel (EU AI Act Art. 53 / US / India DPDP)
• remediation plan: contact templates, license-grant requests, substitute datasets

Flat rate for dataset owners; custom quotes for enterprises. Start the conversation by opening a discussion on this repo, or drop a message on my profile (https://huggingface.co/hardik90).

License

The matrix is factual, API-derived metadata; released under CC0. Verify anything you rely on.