Hugging Face's logo

Datasets:
MJ141592
/
multiple-search-tier1

Tasks:
Question Answering
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

MultipleSearch Tier 1 (Full Phase-1 Build)

Origin

  • Queries: sewon/ambig_qa (train + validation), parsed into ambiguous/unambiguous QA records.
  • Passage source: wiki_snippets (wiki40b_en_100_0) streamed from Hugging Face Datasets.
  • Embeddings: Qwen/Qwen3-Embedding-0.6B.

Processing Summary

  1. Parse AmbigQA records and answer clusters.
  2. Sample 1,000,000 distractor passages (reservoir sampling over wiki stream).
  3. Collect gold passages by:
    • matching article titles from AmbigQA metadata (nq_doc_title / viewed_doc_titles), then
    • matching normalized answer strings in passage text (Aho-Corasick multi-pattern matching).
  4. Link queries to passage IDs via answer-cluster passage matches.
  5. Keep queries with at least one matched gold cluster.
  6. Build train/test split with approximate 50/50 ambiguous/unambiguous balance and 80/20 split.
  7. Embed passages to 1024-d vectors.

Main Files

  • documents.jsonl
  • document_embeddings.pt
  • queries_train.jsonl
  • queries_test.jsonl
  • dataset_stats.json
  • answer_to_passage_ids.json
Downloads last month
8