danny2507/ragu-webq-contriever-msmarco
RAGU WebQuestions retrieval data This repository contains WebQuestions train/dev/test splits prepared for the RAGU codebase. Each record contains question, answers, q_id, and ctxs. Retrieval provenance Retriever: facebook/contriever-msmarco Similarity: exact flat inner product Knowledge base: DPR 2018 English Wikipedia, 100-word passages Retrieved contexts retained per question: 20 WebQ source: official DPR resources Retrieval was performed shard by shard to… See the full description on the dataset page: https://huggingface.co/datasets/danny2507/ragu-webq-contriever-msmarco.
RAGU WebQuestions retrieval data
This repository contains WebQuestions train/dev/test splits prepared for the RAGU codebase. Each record contains question, answers, q_id, and ctxs.
Retrieval provenance
- Retriever:
facebook/contriever-msmarco - Similarity: exact flat inner product
- Knowledge base: DPR 2018 English Wikipedia, 100-word passages
- Retrieved contexts retained per question: 20
- WebQ source: official DPR resources
Retrieval was performed shard by shard to reduce peak RAM. Shard-local top-k results were merged into global top-k results; this is equivalent to searching one flat index over the union of the same embeddings.
Splits
- train: 2,474 questions
- dev: 361 questions
- test: 2,032 questions
The optional webq-test-400-seed10 file follows the RAGU repository's default sampling code. The RAGU paper states that 400 test examples were sampled but does not publish the exact IDs, so this file should not be claimed as the authors' original sample without confirmation.
Important
This is a derived research artifact. Users should review and comply with the licenses and terms of the original WebQuestions, DPR Wikipedia, and Contriever resources. The hasanswer field uses normalized token-sequence matching and is metadata; RAGU's utility training is based on target-model generations, NLI, and Qwen accuracy judgments.