Multilingual Indic RAG Suite
AI4BharatThe Multilingual Indic RAG Suite is a synthetic question-answering dataset of about 21.4 million examples in 17 Indian languages and English, built for retrieval-augmented generation. Llama 3.3 70B wrote each question, answer and reasoning trace from a Wikipedia paragraph, which is kept in the row with its title and URL. AI4Bharat publishes it on Hugging Face as one configuration per language.
Openness
5 medium confidence- license
- mit (card metadata and API tag; the rows carry Wikipedia paragraphs, which Wikipedia publishes under CC BY-SA, and answers written by Llama 3.3 70B)
- access
- public (ungated on the Hub)
- dataset_card
- present (card lists the languages, fields, sample count and the generator model; no paper)
The suite downloads without a gate under MIT, and its card says what each row holds and which model wrote it. The MIT label sits over Wikipedia text, whose own license asks for attribution and share-alike, so a reuser should treat the paragraphs as Wikipedia content rather than as MIT-licensed.
- https://huggingface.co/api/datasets/ai4bharat/Indic-Rag-Suite recorded 2026-11-17
"gated":false; tag license:mit; tags for 18 languages from language:as to language:ur.
- https://huggingface.co/datasets/ai4bharat/Indic-Rag-Suite/raw/main/README.md recorded 2026-11-17
"license: mit"; "Total Samples: 21,439,886"; "Source: Wikipedia + Llama 3.3 70B"; fields "paragraph: Source Wikipedia paragraph", "wiki_id", "url".
Adoption
1 high confidenceHugging Face downloads of the one repository over the trailing month. Users who load a single language configuration are counted the same as those who load all of them.
- https://huggingface.co/api/datasets/ai4bharat/Indic-Rag-Suite recorded 2026-11-17
"downloads":454 for ai4bharat/Indic-Rag-Suite.
Capability
2 medium confidenceThe suite is large and covers low-resource languages such as Santali, Manipuri and Kashmiri, but every question and answer is model-written and nothing beyond its card describes how they were made or checked. Updesh is judged the same way, for the same reason.
- https://huggingface.co/api/models?filter=dataset:ai4bharat/Indic-Rag-Suite recorded 2026-11-17
Two models list the dataset: jangedoo/all-MiniLM-L6-v3-nepali-rag and karmx/Llama-Karmx-Indic-Embedding-bge-m3.
- https://huggingface.co/datasets/ai4bharat/Indic-Rag-Suite/raw/main/README.md recorded 2026-11-17
"A comprehensive multilingual question-answering dataset covering 18 Indian languages with 21,439,886 total samples"; "Source: Wikipedia + Llama 3.3 70B"; "Task: Question Answering with Reasoning".
Verified 2026-11-17, except some scores, still awaiting re-confirmation