witrin/coredd-corpus

CoReDD Corpus

The repository and review context CoReDD Bench is built on.

Overview

CoReDD Corpus holds bare Git histories for six Python projects together with the raw, timestamped pull request, issue, comment, and review events that surrounded them. It exists so that a benchmark run stays reproducible when the upstream repositories change, move, or disappear.

Every event carries a timestamp, which is what lets review context be restricted to whatever existed before a case's cutoff. Without that restriction the ex ante setting of the benchmark cannot be reconstructed.

Nothing here is ground truth. The labels, the uncertainty model, and the evaluation code live in CoReDD Bench, which reads this corpus as its frozen source.

Contents

repositories/
  <owner>/<repository>.tar.gz
issues/
  <owner>/<repository>/<number>.jsonl
  events.parquet
  commits.parquet
  README.md
licenses/
  <owner>/<repository>/
LICENSES.md

• repositories/ contains six complete bare Git repositories packaged as compressed archives.
• issues/ contains 5,689 PR and issue event streams with 155,830 timestamped events. The JSON payloads are raw GitHub REST API fragments. See issues/README.md for the event schema and temporal-use contract.
• events.parquet and commits.parquet provide tabular indexes over the archived material.

Included repositories

• OpenJobDescription/openjd-model-for-python
• mistralai/mistral-common
• cohere-ai/cohere-compass-sdk
• microsoft/Olive
• kedro-org/kedro-plugins
• awslabs/graphstorm

Related resource

CoReDD Bench contains the benchmark labels, audit-derived uncertainty model, and evaluation software. CoReDD Bench uses this corpus as its frozen repository and review context.

Provenance and licensing

All repository and GitHub event material originated from the public upstream projects listed above. Source identity is retained through repository names, Git commit IDs, issue numbers, timestamps, and API payloads.

This is a mixed-rights collection. The repository archives remain subject to their upstream licenses. GitHub issue, PR, comment, and review contents are third-party material and are not relicensed under a blanket dataset license. The CC BY 4.0 license applies only to original CoReDD documentation and corpus-specific material explicitly identified as such. See LICENSES.md.

Removal requests

Requests concerning rights, privacy, or removal can be raised in the Community tab of this dataset. Discussions there are public.

For a non-public or formal route, Hugging Face operates its own reporting and DMCA process under its Content Policy, which applies to this dataset as it does to any repository on the Hub.

Material removed on request is dropped in a new release. Earlier releases are not rewritten, so a benchmark run remains reproducible against the release it was run on.