IN22
AI4BharatIN22 is a machine translation test set for English and all 22 scheduled Indian languages, n-way parallel so that any pair of its languages can be evaluated. IN22-Gen holds 1,024 sentences from Wikipedia and Indian web sources across 13 domains, and IN22-Conv holds 1,503 sentences from English conversations written for the set. In-house translators translated every sentence without machine translation. AI4Bharat built it with the BPCC corpus and the IndicTrans2 models.
Openness
3 high confidence- license
- cc-by-4.0 (card metadata and the IndicTrans2 README's license table)
- access
- auto (Hugging Face gate that approves anyone who agrees to share their contact information)
- dataset_card
- present (cards give domains, fields and sentence counts and point to the dataset card appendix of the paper)
The sentences carry CC BY 4.0, and the Hugging Face copies need an access request that is approved automatically. The cards and the paper describe how the sentences were chosen and translated.
- https://huggingface.co/api/datasets/ai4bharat/IN22-Conv recorded 2026-11-17
API JSON: "gated":"auto"; card metadata "license":"cc-by-4.0", split "test".
- https://huggingface.co/api/datasets/ai4bharat/IN22-Gen recorded 2026-11-17
API JSON: "gated":"auto"; card metadata "license":"cc-by-4.0", split "test".
- https://huggingface.co/datasets/ai4bharat/IN22-Conv recorded 2026-11-17
Card: "You need to agree to share your Hugging Face information to access this dataset"; "License: cc-by-4.0"; "The evaluation subset consists of 1503 sentences translated across 22 Indic languages".
- https://huggingface.co/datasets/ai4bharat/IN22-Gen recorded 2026-11-17
Card: "You need to agree to share your Hugging Face information to access this dataset"; "License: cc-by-4.0"; "The evaluation subset consists of 1024 sentences translated across 22 Indic languages"; "Please refer to the Appendix E: Dataset Card of the preprint".
- https://raw.githubusercontent.com/AI4Bharat/IndicTrans2/main/README.md recorded 2026-11-17
License table: "Newly Created IN-22 test set (IN22-Gen & IN22-Conv) | CC-BY-4.0"; "IN22-Gen (IN22-Wiki + IN22-Web) | 1024 sentences"; "IN22-Conv | 1503 sentences".
Adoption
2 high confidenceHugging Face downloads of the IN22-Gen and IN22-Conv repositories over the trailing month. Evaluation harnesses can fetch a test set repeatedly, so the figure is a rough measure of use, not a count of users; copies taken from the archive in the BPCC repository are not counted.
- https://huggingface.co/api/datasets/ai4bharat/IN22-Conv recorded 2026-11-17
"downloads":1123 for ai4bharat/IN22-Conv.
- https://huggingface.co/api/datasets/ai4bharat/IN22-Gen recorded 2026-11-17
"downloads":1495 for ai4bharat/IN22-Gen.
Capability
3 medium confidenceIn every language IN22 targets, its deeper test, the human-translated conversation sentences of IN22-Conv, has four hundred to sixteen hundred items that count toward depth. An item count says how much a set can test, not how precisely it ranks models.
- https://arxiv.org/abs/2305.16307 recorded 2026-11-17
Abstract: "Our second contribution is the release of the first n-way parallel benchmark covering all 22 Indian languages, featuring diverse domains, Indian-origin content, and source-original test sets."
- https://arxiv.org/pdf/2305.16307 recorded 2026-11-17
"It is an n-way parallel test set containing 2527 original English sentences translated into 22 Indic languages with high-quality translations done by in-house translators from scratch without recourse to any existing MT system"; Table 2: "IN22-Wiki 512 22", "IN22-Web 512 22", "IN22-Conv 1503 22"; "we merge Wikipedia and Web Sources subsets, to create a consolidated set referred to as IN22-Gen".
- https://huggingface.co/datasets/ai4bharat/IN22-Conv recorded 2026-11-17
"The evaluation subset consists of 1503 sentences translated across 22 Indic languages"; split "test".
- https://huggingface.co/datasets/ai4bharat/IN22-Gen recorded 2026-11-17
"The evaluation subset consists of 1024 sentences translated across 22 Indic languages"; split "test".
Verified 2026-11-17, except some scores, still awaiting re-confirmation