65% of Indic-language datasets on the Hub declare no license — full matrix published

#1
by hardik90 - opened

I scanned 4,893 Indic-language datasets (hindi, tamil, bengali, telugu, marathi, malayalam, kannada, gujarati, punjabi, urdu, odia, assamese, nepali, hinglish, code-mixed + more) and published the full results here:

https://huggingface.co/datasets/hardik90/indic-dataset-license-matrix

Headline numbers (scan date 2026-08-01):

  • 3,185 of 4,893 datasets (65.1%) declare no license at all
  • 46.1% of all downloads hit no-license repos — including ai4bharat/IndicCorpV2 and cfilt/iitb-english-hindi
  • 7.7% of downloads hit non-commercial (CC-BY-NC*) repos — a silent trap for anyone training a commercial product (e.g. ai4bharat/IndicParaphrase, 4,611 downloads)
  • 139 datasets carry ambiguous tags (other, cc, unknown)

Why this matters: under the EU AI Act GPAI training-data documentation duties and most enterprise open-source policies, a missing license tag effectively disqualifies a dataset from commercial pipelines — regardless of what the original authors intended.

The matrix includes a risk bucket + one-line guidance for every repo, so dataset owners can find and fix their own entries, and companies can build a watchlist before training.

@ai4bharat @cfilt @krutrim-ai-labs — several flagship corpora appear in the high-risk list; I'd love your maintainers to add license tags + provenance notes so adoption keeps growing.

The CSV is CC0 (factual API metadata). Feedback and corrections welcome — I'll re-run the scan monthly.

— Hardik

Sign up or log in to comment