Datasets:

Hyukkyu
/

bright-stackoverflow

pretty_name: BRIGHT stackoverflow (unified retrieval schema)
license: cc-by-4.0
language:
  - en
task_categories:
  - text-retrieval
tags:
  - bright
  - retrieval
  - reasoning
configs:
  - config_name: corpus
    data_files:
      - split: test
        path: corpus/test-*.parquet
  - config_name: queries
    data_files:
      - split: test
        path: queries/test-*.parquet
  - config_name: qrels
    data_files:
      - split: test
        path: qrels/test-*.parquet

BRIGHT: stackoverflow

A format conversion of the standard passage-retrieval setting of BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval, by the original BRIGHT authors. Source: xlangai/BRIGHT, pinned to 3066d29c9651a576c8aba4832d249807b181ecae. This is a converted mirror, not an independently annotated dataset. The original dataset is distributed under CC BY 4.0. See SOURCE_README.md for the original dataset card and attribution.

Table	Rows	Columns
corpus	107,081	id, title, text
queries	117	id, text
qrels	478	query-id, corpus-id, score
exclusions (optional)	0	query-id, corpus-id

All published configs use the test split. corpus.text is exactly the original documents.content; title is empty. Query IDs, document IDs and original examples.query strings are preserved without normalization. qrels is the unique set of gold_ids pairs, with binary score 1. Duplicate labels are collapsed; no relevance judgments are added. Gold reasoning, gold answers, gold_ids_long, generated query rewrites and long_documents are not part of the converted evaluation inputs.

This subset has no real excluded documents; the optional exclusions table is omitted. The N/A exclusion sentinel is removed. Excluded IDs and positive IDs are validated against the corpus and must be disjoint for each query. Keep ignore_identical_ids=false: matching query/document ID strings alone do not imply a BRIGHT exclusion.

Evaluate each BRIGHT subset against its own complete corpus. Report nDCG@10 per subset and the unweighted mean across the 12 subsets. Record encoder context limits and truncation. The original evaluation code defines the protocol.

Converted with scripts/preprocess/dataset/convert_bright.py in the SPARSE project. conversion.json records source and output SHA-256 hashes, row counts and validation checks.