BharatGen Datasets
BharatGen’s datasets are built to fill real gaps: benchmarks for evaluating AI on Indian languages, domains and knowledge systems that are underrepresented or entirely absent from global evaluation suites.
Our datasets cover legal, financial, agricultural and Ayurvedic knowledge, mental health question answering from medical literature, low and extremely low-resource Indic language benchmarks from official examination corpora and more. All datasets are openly available for research.
| No. | Dataset | Published by | Date | Download on Hugging Face | View on AI Kosh |
|---|---|---|---|---|---|
| 1. | IndicParam A graduate-level MCQ benchmark for evaluating LLMs on 11 low and extremely low-resource Indic languages, sourced from UGC-NET exam papers. | IIM Indore | Dec 2025 | ↗ | — |
| 2. | ParamBench A graduate-level MCQ benchmark for evaluating LLM understanding of Indic subjects in Hindi, spanning 21 domains from competitive examinations. | IIM Indore | Nov 2025 | ↗ | — |
| 3. | TORQUE A Hindi benchmark for table reconstruction and visual QA in Devanagari script, covering 210 table images from government documents and printed sources. | BharatGen | Nov 2025 | ↗ | — |
| 4. | BhashaBench-Legal 24,365 expert-validated questions on Indian legal knowledge in English and Hindi, sourced from 50+ judicial service and bar examinations across 20+ legal domains. | BharatGen | Oct 2025 | ↗ | ↗ |
| 5. | BhashaBench-Finance 19,433 questions on Indian financial knowledge in English and Hindi, sourced from 25+ government financial examinations across 30+ domains. | BharatGen | Oct 2025 | ↗ | ↗ |
| 6. | BhashaBench-Krishi15,405 questions on Indian agricultural knowledge in English and Hindi, sourced from 55+ government agricultural examinations across 25+ domains. | BharatGen | Oct 2025 | ↗ | ↗ |
| 7. | BhashaBench-Ayur14,963 questions on Ayurvedic knowledge in English and Hindi, sourced from 50+ authentic government Ayurvedic examinations across 15+ disciplines. | BharatGen | Oct 2025 | ↗ | ↗ |
| 8. | S3VQA: Select, Substitute, and SearchAn open-domain visual question answering dataset using a reformulate-then-retrieve approach for knowledge-grounded VQA. | IIT Bombay | May 2025 | — | ↗ |
| 9. | DictDisA dataset for lexically constrained neural machine translation using domain-specific dictionaries with multiple candidate translations per source word. | IIT Bombay | May 2025 | — | ↗ |
| 10. | RUDDER: Cross-Lingual Video and Text RetrievalVideos of scientific toy-making with paired sentences and audio in multiple languages, for cross-lingual video-text retrieval. | IIT Bombay | May 2025 | — | ↗ |
| 11. | CATALIST: Camera Transformations for Multi-Lingual Scene Text RecognitionScene-text videos with camera movement annotations and word-level masks for multilingual OCR in natural settings. | IIT Bombay | May 2025 | — | ↗ |
| 12. | IndicQA: Multilingual Question Answering BenchmarkContext-grounded question-answer pairs across 11 Indian languages for evaluating multilingual question-answering systems. | IIT Bombay | May 2025 | — | ↗ |
| 13. | BharatGen MHQA Dataset2,475 expert-verified mental health QA pairs covering anxiety, depression, trauma, and OCD, built from PubMed abstracts across factoid, diagnostic, prognostic, and preventive question types. | BharatGen | Apr 2025 | — | ↗ |
| 14. | AVVP: Audio Visual Video ParsingDetects temporal boundaries of audio-visual events by modelling how event categories bind to audio and visual modalities. | IIT Bombay | Apr 2025 | — | ↗ |
| 15. | Post-OCR Text Correction in Sanskrit218,000 sentences across 30 books for benchmarking OCR correction in Sanskrit, spanning domains from astronomy to medicine. | IIT Bombay | Apr 2025 | — | ↗ |
| 16. | MALTA: Multi-Modal and Multi-Lingual Temporal Sentence AlignmentVideo tutorials in English and Marathi featuring toy-making from waste material and organic farming demonstrations. | IIT Bombay | Apr 2025 | — | ↗ |
| 17. | Math Word Problems in Hindi and EnglishA weakly supervised dataset for solving math word problems in Hindi and English using only final answers as supervision signals. | IIT Bombay | Apr 2025 | — | ↗ |
| 18. | MTD-360: Multilingual Table Detection360 document images spanning 14 languages and diverse formats (textbooks, bank reports, bills) for benchmarking table detection models. | IIT Bombay | Apr 2025 | — | ↗ |
| 19. | Saamayik: English-Sanskrit Parallel Corpus~53,000 parallel English-Sanskrit sentences in contemporary prose, spanning language instruction, pedagogy, and online tutorials. | IIT Bombay | Apr 2025 | — | ↗ |
| 20. | Vāksañcayaḥ: Sanskrit ASR Corpus78+ hours of Sanskrit speech covering 45,953 sentences from Śāstras, contemporary stories, radio programs, and extempore discourse. | IIT Bombay | Apr 2025 | — | ↗ |
| 21. | Dharmawiki: Indic Knowledge CorpusA multilingual collection of Indian philosophical, religious, and cultural content for developing and evaluating Indic language technologies. | IIT Bombay | Apr 2025 | — | ↗ |