Datasets:

MapleBi
/

MetaRAG_Cross-Issue_OSSQA

language:
  - en
license: other
license_name: github-terms-of-service
task_categories:
  - question-answering
  - information-retrieval
tags:
  - rag
  - github-issues
  - cross-issue-reasoning
  - silver-labels
pretty_name: MetaRAG Cross-Issue OSSQA
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    default: true
    data_files:
      - split: dev
        path: data/default/dev.jsonl
      - split: test
        path: data/default/test.jsonl
  - config_name: queries
    data_files:
      - split: dev
        path: data/queries/dev.jsonl
      - split: test
        path: data/queries/test.jsonl
  - config_name: corpus
    data_files:
      - split: corpus
        path: data/corpus/corpus.jsonl

MetaRAG Cross-Issue OSSQA

Dataset page: https://huggingface.co/datasets/MapleBi/MetaRAG_Cross-Issue_OSSQA

MetaRAG Cross-Issue OSSQA is an English open-source software issue question-answering and retrieval benchmark. Each example asks a question grounded in one GitHub issue and requires evidence from a related issue. The data contains explicit cross-issue references and a three-document silver evidence path.

Dataset configurations

Configuration	Splits	Rows	Purpose
`default`	`dev`, `test`	100 / 400	Complete evaluation records, including embedded issue fragments and silver labels.
`queries`	`dev`, `test`	100 / 400	Retrieval and QA queries with answer and evidence-path supervision.
`corpus`	`corpus`	9,536	Deduplicated issue-description and issue-comment fragments for retrieval. This is a knowledge corpus, not a training split.

Loading

from datasets import load_dataset

evaluation = load_dataset("MapleBi/MetaRAG_Cross-Issue_OSSQA")
queries = load_dataset("MapleBi/MetaRAG_Cross-Issue_OSSQA", "queries")
corpus = load_dataset("MapleBi/MetaRAG_Cross-Issue_OSSQA", "corpus", split="corpus")

Main fields

default includes the query, its known context, answer, source and target issue metadata, silver_evidence_path, quality signals, and the ordered fragments supporting each example. queries is the compact query-level representation. corpus contains one searchable fragment per document_id, plus issue, repository, author, timestamp, and cross-issue-link metadata.

The fields answer_document_id and silver_evidence_path refer to corpus.document_id. known_context_document_ids identifies documents supplied as query context; these are excluded from the gold evidence path. split is repository-disjoint between dev and test.

Data creation

The dataset was constructed from public GitHub event data via OpenDigger / GH Archive, from 2022-01-01 through 2025-12-31. It contains 500 qualified cross-issue samples from 468 repositories. The sampling procedure requires an explicit cross-issue reference, a path spanning at least two distinct issues, English-language filtering, deterministic diversity sampling, and repository-disjoint development/test partitions.

See dataset_statistics.json for the complete sampling settings, quality distribution, year distribution, and leakage audit result.

Labels and limitations

reference_answer and silver_evidence_path are heuristic silver annotations. They should not be treated as human-validated gold labels without additional review. Source text may contain stale links, personal identifiers visible in public GitHub content, inaccurate technical claims, or other artifacts of public issue discussions.

The data is intended for research and evaluation. Users are responsible for ensuring their use complies with the applicable GitHub Terms of Service, the source platforms' policies, and their own privacy, legal, and ethical requirements.

Provenance

OpenDigger: https://github.com/X-lab2017/open-digger
GH Archive: https://www.gharchive.org/

Repository integrity

MANIFEST.sha256 contains SHA-256 checksums for every data file and dataset_statistics.json in this package. To regenerate the package from the accompanying source files, remove this output directory and run python3 prepare_hf_dataset.py.