bharath.ai › Made in India
Models and datasets from Indian labs, with what each reported and a link to the paper or the lab’s own page. A result a lab reports about its own model is marked as its own claim.
ARTPARK and Dept. of Electrical Engineering, IISc Bangalore (with Google DeepMind)
31,255 hours of speech, 2,043 hours of it transcribed, and 289,838 images across 105 languages and 165 districts (Table 1).
Read it ↗Sarvam AI, trained in India on IndiaAI Mission compute
The company’s own benchmark page reports MMLU 90.6, GPQA Diamond 78.7 and Math500 98.6 for Sarvam 105B. These are self-reported by the lab, not an independent evaluation.
Read it ↗Krutrim AI, Bangalore
Scores 67.90 on English and 51.5 on Hindi in BharatBench LLaVA-Bench, and 87.63 on POPE English (Table 1).
Read it ↗Krutrim AI, Bangalore (Ola Krutrim)
Reports 0.89 BERTScore on IndicCOPA Bengali against GPT-3.5’s 0.77 (Table 4).
Read it ↗AI4Bharat, Dept. of CSE, IIT Madras
1,704 hours from 10,496 speakers across all 22 scheduled languages (Table 3). Published at NeurIPS 2024 Datasets and Benchmarks.
Read it ↗AI4Bharat Nilekani Centre, IIT Madras (with Sarvam AI)
The Sangraha corpus totals 251,321 million tokens: 64B verified, 24B unverified and 162B synthetic (Table 1).
Read it ↗AI4Bharat, IIT Madras (with Sarvam AI)
7,348 hours from 16,237 speakers across 145 districts and 22 languages (Table 1).
Read it ↗LTRC, KCIS, IIIT Hyderabad
3.39 million article-headline pairs across 8 languages; the best baseline averages 31.43 ROUGE-L (Table 2).
Read it ↗AI4Bharat Nilekani Centre, IIT Madras
Gains 5.2 chrF++ over IndicTrans1 on English-to-Indic and 6.3 on Indic-to-English, and beats the best commercial and open baselines on English-to-Indic (Table 12, IN22-Gen). Published in TMLR, December 2023.
Read it ↗