Hugging Face's logo

Datasets:
williamTLmiller
/
swe-bench-coding-agent-tasks

Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Sub-tasks:
language-modeling
Languages:
English
Size:
10K - 100K
Tags:
code
code-generation
software-engineering
program-repair
bug-fixing
github
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
swe-bench-coding-agent-tasks / README.md
williamTLmiller's picture
williamTLmiller
Card metadata: full task/size/domain tags and search-friendly title
155f934 verified
preview code
|
Raw History Blame Contribute Delete
6.75 kB
metadata 
pretty_name: >-
  SWE-bench Style Coding Agent Dataset: GitHub Issues, Patches & Tests
  (Decontaminated)
license: cc-by-nc-4.0
language:
  - en
multilinguality:
  - monolingual
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
task_ids:
  - language-modeling
annotations_creators:
  - machine-generated
language_creators:
  - found
source_datasets:
  - original
tags:
  - code
  - code-generation
  - software-engineering
  - program-repair
  - bug-fixing
  - github
  - github-issues
  - pull-requests
  - swe-bench
  - swe-smith
  - coding-agent
  - agents
  - llm-evaluation
  - benchmark
  - contamination
  - decontamination
  - python
  - matrixnorm
configs:
  - config_name: agent_tasks
    data_files:
      - split: train
        path: parquet/agent_tasks.parquet
  - config_name: contamination_groups
    data_files:
      - split: train
        path: parquet/contamination_groups.parquet
  - config_name: diff_hunks
    data_files:
      - split: train
        path: parquet/diff_hunks.parquet
  - config_name: environments
    data_files:
      - split: train
        path: parquet/environments.parquet
  - config_name: eval_tasks
    data_files:
      - split: train
        path: parquet/eval_tasks.parquet
  - config_name: execution_proofs
    data_files:
      - split: train
        path: parquet/execution_proofs.parquet
  - config_name: heldout_eval
    data_files:
      - split: train
        path: parquet/heldout_eval.parquet
  - config_name: identity_groups
    data_files:
      - split: train
        path: parquet/identity_groups.parquet
  - config_name: issues
    data_files:
      - split: train
        path: parquet/issues.parquet
  - config_name: pull_requests
    data_files:
      - split: train
        path: parquet/pull_requests.parquet
  - config_name: quality
    data_files:
      - split: train
        path: parquet/quality.parquet
  - config_name: repos
    data_files:
      - split: train
        path: parquet/repos.parquet
  - config_name: split_manifests
    data_files:
      - split: train
        path: parquet/split_manifests.parquet
  - config_name: split_memberships
    data_files:
      - split: train
        path: parquet/split_memberships.parquet
  - config_name: task_validity
    data_files:
      - split: train
        path: parquet/task_validity.parquet
  - config_name: tests
    data_files:
      - split: train
        path: parquet/tests.parquet
  - config_name: train
    data_files:
      - split: train
        path: parquet/train.parquet

SWE-bench Style Coding Agent Dataset: GitHub Issues, Patches & Tests (Decontaminated)

Release name: SWEGraph v2 — Free Sample

A free, deterministic ~10% sample of SWEGraph v2: the complete SWE-bench family, SWE-smith and BigCodeBench normalized into proof chains (repo -> issue -> patch -> diff hunks -> tests -> agent task -> eval task) with identity/contamination groups and a leakage-audited split. The full package is sold separately (see below).

  • License: CC BY-NC 4.0 for the MatrixNorm normalization/packaging layer (see Licensing).
  • Format: gzipped CSV in csv/ (all columns as text; list fields pipe-delimited) plus a Parquet mirror in parquet/ with identical rows, used by load_dataset and the viewer.
from datasets import load_dataset
tasks = load_dataset("williamTLmiller/matrixnorm-swegraph-v2-sample", "agent_tasks", split="train")

How the sample was drawn

  • Unit = repository family (a SWE-smith mirror belongs to its upstream's family). Keep a family iff the first byte of SHA-1("sample:" + family name) is < 26 (of 256). Families are kept whole, so contamination groups and partitions are never cut.
  • Same rule on every table: tables with a repo_id keep the sampled families; group, quality, validity, proof, environment and membership rows are kept when they refer to a kept task or record. split_manifests is included in full.
  • The split column is the full package's partition, unchanged. Because whole families are sampled, a sample may contain no validation or test family.
  • Built 2026-10-23T22:50:54Z.

Rows per table

Table Full package This sample
agent_tasks 83,357 6,265
contamination_groups 250 23
diff_hunks 136,336 11,892
environments 642 128
eval_tasks 5,213 353
execution_proofs 5,213 353
heldout_eval 16,822 1,663
identity_groups 21,073 1,260
issues 82,217 6,265
pull_requests 82,217 6,265
quality 254,489 19,173
repos 345 25
split_manifests 28 28
split_memberships 88,570 6,618
symbols 1,140 0
task_validity 5,213 353
test 764 0
tests 235,416 16,393
train 70,602 4,955
validation 382 0

The full dataset

The licensed package holds every table above in full, plus per-entity JSONL streams with the complete nested fields (including each task's full PASS_TO_PASS regression list), provenance, rights and leak-audit files. Perpetual, non-exclusive license for internal research, training and benchmarking; models you train are yours to commercialize; no redistribution of the dataset itself. Contact: energea@energeatech.com

Sources

Source Rows used License of the source release
SWE-bench (test, dev, train) 2,294 + 225 + 19,008 MIT
SWE-bench Verified / Lite / Multimodal 500 / 323 / 612 (subsets and extensions of SWE-bench) MIT
SWE-bench Multilingual 300 MIT
SWE-bench Pro V2 (default; hard recorded as membership) 642 MIT tooling; task content under upstream repo licenses
SWE-smith 59,136 MIT
BigCodeBench v0.1.4 1,140 Apache-2.0

Task content (issue text, patches, tests) comes from public open-source repositories and remains under each repository's own license; the exporter only ships verbatim code text where rights_manifest.csv allows it.

Licensing

  • This sample: CC BY-NC 4.0 — attribution, non-commercial — applies to the MatrixNorm normalization and packaging layer.
  • Underlying data: the source releases keep their own licenses (copies included as LICENSE-*-UNDERLYING-*.txt); repository content stays under each repository's license. See NOTICE.
  • Commercial use of the normalized package requires the paid license.

Limitations

  • Task validity gates are not_run until real execution evidence exists.
  • Issue and PR text is reproduced from public GitHub as published in the source releases.
  • Not produced, reviewed or endorsed by the authors of any source dataset.

Citation

SWEGraph v2 (free sample). MatrixNorm normalization by williamTLmiller, derived from SWE-bench, SWE-smith, SWE-bench Pro and BigCodeBench. https://huggingface.co/datasets/williamTLmiller/matrixnorm-swegraph-v2-sample

Contact

energea@energeatech.com