Indic Dataset Quality Index: 4,893 repos scored — junk, duplicates and template-card spam found
#1
by hardik90 - opened
I scored all 4,893 Indic-language datasets on the Hub for quality (12 automated signals: junk names, boilerplate cards, duplicates, dead leaderboard artifacts, staleness) and published the full index:
https://huggingface.co/datasets/hardik90/indic-dataset-quality-index
Headline numbers:
- GOLD 1,904 (38.9%) / OK 2,969 / DEGRADED 16 / JUNK 4
- 113 duplicate clusters — e.g. cfilt/iitb-english-hindi re-published as chungimungi/iitb-english-hindi-split; aaditya/orca_dpo_pairs-Hindi vs ...-Hindi_ (same corpus, two repos)
- 2 dead leaderboard artifacts with 1,000+ downloads (open-llm-leaderboard-old/details_Telugu-LLM-Labs__*) — cards that look like real datasets
- 2 template-card spam repos (unchanged HF boilerplate)
- A 23-repo cluster sharing one boilerplate card (cmeraki/ultrachat_hindi_seamless + 22 others)
- The saillab/alpaca_*_taco family: 13 language variants, identical card text — verify content before trusting
Per-repo CSV + Parquet with quality score (0-100), bucket, and every signal that fired. Companion dataset: the License Matrix (65% no-license — the bigger structural risk).
If you maintain any flagged repo: I offer a quality verification + repair service (fixes, card rewrite, verification badge) — flat rate for owners, custom for teams. Open a discussion or message my profile.
— Hardik