williamTLmiller/swe-bench-coding-agent-tasks

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

Release name: SWEGraph v3 — Free Toy Sample

A free, deterministic, deliberately tiny toy sample of SWEGraph v3: 950 agent tasks from 25 repository families, with every row linked to them (50 of the tasks also appear as eval tasks). SWEGraph v3 is the complete SWE-bench family, SWE-smith and BigCodeBench normalized into proof chains (repo -> issue -> fix -> diff hunks -> tests -> agent task -> eval task), each task linked to its license, its container image and its execution evidence, with identity/contamination groups and a leakage-audited split. The full package is sold separately (see below).

Free toy sample

This public repo is a deliberately tiny, CSV-only toy sample: 950 agent tasks plus the rows linked to them. It is meant for looking at the schema and running a pipeline once (a toy training run), not for training a real model. The full and training-ready versions (williamTLmiller/swe-bench-coding-agent-tasks-full) are private and licensed separately; contact energea@energeatech.com.

• License: CC BY-NC 4.0 for the MatrixNorm normalization/packaging layer (see Licensing).
• Format: plain CSV in csv/ plus repo_licenses.csv at the root (list fields pipe-delimited, dicts as JSON). Every data file in this repo is a plain CSV table.
• Metadata only: the text columns (problem_statement, hints_text, requirements, interface, patch, test_patch, bug_patch, solution, title, input_summary, the issue body and diff lines) are empty in this sample. Licenses, images, difficulty, tests, gate results and validity labels are all here. The full package carries the text for every permissively licensed task.

from datasets import load_dataset
tasks = load_dataset("williamTLmiller/swe-bench-coding-agent-tasks", "agent_tasks", split="train")

What is in this toy sample

Source	split	Agent tasks
SWE-bench train	train	420
SWE-smith	train	380
SWE-smith	heldout	100
SWE-bench Multimodal	heldout	25
SWE-bench Pro V2	heldout	25
Total		950

• Split: 800 train and 150 heldout tasks (the full package's partition, unchanged). The 50 SWE-bench Multimodal and Pro tasks are also in eval_tasks.csv as eval tasks, so heldout_eval.csv has 200 rows (100 SWE-smith tasks + 50 tasks x 2 views); train.csv holds the 800 train tasks.
• Difficulty tier: easy 330, medium 366, hard 162, very hard 92.
• Code text status (what the full package ships): included (permissive license) 922, withheld (copyleft) 25, withheld (license not detected)
• Execution evidence: 530 tasks carry what the source's authors executed (upstream_reported); 420 SWE-bench train tasks have a patch and tests only (patch_only). 530 tasks are linked to a published container image.
• Languages: Python 859, Go 57, JavaScript 34.
• Tests: 3,398 FAIL_TO_PASS and 216 PASS_TO_PASS test rows, the tests named by the sampled tasks.
• Issue dates (created_at, SWE-bench train issues): 2014-05-02 to 2023-08-21.

Safe task setup

docker run -it --network none <environment_image> bash
cd /testbed                       # /app for SWE-bench Pro
git apply bug.diff                # SWE-smith only: bug_patch sets up the task's start state
rm -rf .git && git init -q && git add -A && git -c user.name=t -c user.email=t@t commit -qm task

Re-initialising git removes every ref, reflog entry and object an agent could read the answer from (git log --all, other tasks' bug branches, upstream commits made after the base commit). Do it for every task, from every source, before the agent starts.

How the toy sample was drawn

• Unit = agent task, drawn from the 25 repository families of the earlier, larger free sample. Tasks are stratified by source and split (100 SWE-smith heldout, 25 SWE-bench Multimodal, 25 SWE-bench Pro, 380 SWE-smith train, 420 SWE-bench train) and, inside each stratum, taken in a fixed salted-SHA-256 order of the instance id (sha256("matrixnorm-toy-v1|" + instance_id)), so the pick is deterministic.
• Tasks whose FAIL_TO_PASS or test-id list is longer than 32,767 characters (a spreadsheet cell's limit) are left out of the toy; they are in the full package.
• Linked rows follow the tasks: issues, pull requests, diff hunks, tests, eval tasks, execution proofs, task validity, quality, split memberships, environments and repositories keep the rows of the sampled tasks; train.csv / heldout_eval.csv are cut to them. split_manifests (28 rows) is included in full, so its partition_counts and group_count are full-package counts.
• Groups: identity and contamination groups are kept when a sampled task, issue, pull request or repository belongs to them, and their member lists are cut to the members in this sample (member_count counts those); canonical_id is unchanged and can name a member that is only in the full package. Like the full package, the identity table stores only groups with more than one member, so an id with no row is a one-member group.
• repo_licenses.csv covers the 25 sampled repositories (SWE-smith mirrors under their upstream name); its task_rows counts this sample's agent tasks.

Rows per table

File	Full package	This toy sample
csv/agent_tasks.csv	83,357	950
csv/contamination_groups.csv	248	23
csv/diff_hunks.csv	140,106	2,067
csv/environments.csv	4,295	65
csv/eval_tasks.csv	5,213	50
csv/execution_proofs.csv	83,357	950
csv/heldout_eval.csv	17,027	200
csv/identity_groups.csv	21,294	363
csv/issues.csv	82,217	950
csv/pull_requests.csv	82,217	950
csv/quality.csv	254,489	2,925
csv/repos.csv	345	25
csv/split_manifests.csv	28	28
csv/split_memberships.csv	88,570	1,000
csv/task_validity.csv	83,357	950
csv/tests.csv	235,416	3,614
csv/train.csv	70,017	800
repo_licenses.csv	334	25

The full package also has symbols, validation and test tables and validation_runs.csv (none of their rows belong to the sampled tasks).

About the full package

The sections below describe the full SWEGraph v3 package; every count in them is a full-package count, not a count of this toy sample.

What's new in v3

v2 shipped metadata only: every repository's license was unknown, so the issue text, patches and tests were blanked, no task was linked to a runnable image, and nothing had been executed. v3 fixes that:

• Licenses for 334 repositories: 331 from the LICENSE/COPYING file at probed base commits (each repository's earliest and latest base commit, plus SWE-smith pinned commits); 2 from deps.dev's current default-branch license because no license file classified (hamba/avro, tokio-rs/axum). deps.dev reports the same license for 255 of them. Text now ships for 77,921 of 83,357 tasks (93%) whose repository is permissively licensed; copyleft, source-available and undetected licenses keep text withheld.
• SWE-smith fixes corrected. SWE-smith's patch column is the diff that introduces the bug. v2 stored it as the fix. v3 ships the exact reverse as patch (the gold fix) and the original as bug_patch (the diff that sets up the task) for its 59,136 SWE-smith tasks. In every one of the 21 SWE-smith container runs the reversed fix restored the clean code byte for byte.
• 63,160 of 83,357 tasks linked to their published container image with digest and size. The rest have none (image not found in the registry: SWE-bench test/dev 49; no per-task image published: SWE-bench train 19,008, BigCodeBench 1,140).
• Real execution evidence on a stratified sample: 49 tasks run by us in their containers (gold fix twice, a no-op patch, and a git-history audit inside the image), 48 of which produced test results; 63,131 carry what the source's authors executed, marked upstream_reported, never as our own pass.
• A difficulty tier on every task, by basis: gold-fix size 81,160, not_labelled_by_source 1,140, no public gold patch to size (unknown) 506, SWE-bench Verified's human labels 500, swebench_pro_hard_subset+gold_patch_size 39, Pro's hard subset 12. Plus error_message, stack_trace, expected_behavior, actual_behavior and reproduction_steps pulled out of the issues, repository stars/forks/open issues, language and test framework.

Licenses and text

Value	Tasks	Share
included (permissive license)	77,921	93.5%
withheld: copyleft (GPL, LGPL, AGPL, MPL)	4,935	5.9%
withheld: license not detected	406	0.5%
withheld: source-available (BUSL)	94	0.1%
withheld: text clashes with a reserved field name	1	0.0%

The full release's repo_licenses.csv lists each repository's license, the file it was read from, every license seen across its probed commits (licenses_seen) and whether it changed between them (it did for 9: PrefectHQ/prefect (Apache-2.0, none detected at some commits); conda/conda (BSD-3-Clause, none detected at some commits); dagster-io/dagster (Apache-2.0, none detected at some commits); docker/compose (Apache-2.0, BSD-3-Clause); future-architect/vuls (AGPL-3.0, GPL-3.0); psf/requests (Apache-2.0, ISC); redis/redis (BSD-3-Clause, SSPL-1.0); sphinx-doc/sphinx (BSD-2-Clause, BSD-2-Clause AND BSD-3-Clause); ytdl-org/youtube-dl (Unlicense, none detected at some commits)). 61,133 of 83,357 tasks sit at a probed base commit and use that commit's license; every other task uses the repository-level spdx.

Execution evidence

Value	Tasks	Share
the source's authors ran it (their harness)	63,131	75.7%
never executed (patch and tests only)	19,008	22.8%
not executed (unit tests ship with the task)	1,140	1.4%
we ran it in its container	48	0.1%
tests not published by the source	30	0.0%

Our container runs

Official swebench harness for the SWE-bench family; for SWE-smith, the bug patch, then the reversed fix, then the repository profile's test command. Each task: gold fix twice, a no-op patch (tests must fail without the fix), and a git-history audit of the image.

Source	Tasks run	Image runs	Fix passes its tests	Tests fail without the fix	Other tests stay green	Same result twice	No answer in git history
SWE-bench Verified	8	8	8	6	5	8	8
SWE-bench Lite	4	4	3	3	3	4	1
SWE-bench test	4	4	4	4	3	4	0
SWE-bench dev	2	2	2	2	2	2	0
SWE-bench Multilingual	8	8	8	8	8	8	8
SWE-bench Multimodal	2	2	1	1	1	1	2
SWE-smith	17	17	17	17	17	17	0
SWE-smith (files without a final newline)	4	4	4	4	4	4	0

42 of 49 tasks passed all five core checks. Every result, including the failures, is in the full release's validation_runs.csv and on the task (gate_results, validity_status). None of the 49 container-run tasks is in this toy sample.

What the runs found:

• The reversed SWE-smith fix: the bug patch applied cleanly in every one of the 21 SWE-smith runs, and the reversed fix restored the exact clean tree in every one of the 21.
• Answer leak in SWE-bench images: 9 of 28 audited SWE-bench runs (SWE-bench Lite 3 of 4, SWE-bench dev 2 of 2, SWE-bench test 4 of 4) found 397 to 4,257 commits newer than the checked-out commit; the image is checked out at the base commit but keeps the upstream repository's later history in its refs and object store, so later commits (where the real fix lives) can be read from git. SWE-bench images are per task: a SWE-bench task we did not run has git_history_audit = not_run, not clean. Run tasks are flagged git_history_audit = leak_found. Use the safe task setup above for every task.
• Answer leak in SWE-smith images: 21 of 21 audited SWE-smith runs (SWE-smith 17 of 17, SWE-smith (files without a final newline) 4 of 4) found 0 to 1,358 commits newer than the checked-out commit; their git history includes branches for other tasks' bug commits (origin/<instance_id>), so git log --all can show where a bug was injected. Tasks on the same image are flagged leak_found_in_image. Run tasks are flagged git_history_audit = leak_found. Use the safe task setup above for every task.
• Some source labels do not hold today in the published image: the gold fix did not pass its FAIL_TO_PASS tests in 1 run (SWE-bench); FAIL_TO_PASS tests already passed without the fix in 3 runs (SWE-bench); a PASS_TO_PASS test failed with the gold fix in 5 runs (SWE-bench). Those tasks are marked (gate_results, validity_status) instead of hidden.

The full dataset

The licensed package holds every table above in full with the issue, patch and test text, training partitions, per-entity streams with the complete nested fields (including each task's full PASS_TO_PASS 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 in the full package	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
SWE-bench org copies of Verified / Lite / SWE-bench / Multimodal	harness fields only (image, log parser, difficulty) and the Multimodal test patches the princeton-nlp copy leaves empty	MIT

Task content (issue text, patches, tests) comes from public open-source repositories and remains under each repository's own license. Each repository's license was detected from its LICENSE/COPYING file at probed base commits, or from deps.dev's default-branch license where no file classified (repo_licenses.csv gives the evidence per repository); verbatim text ships (in the full package) only for permissive licenses (code_text_status = included).

Licensing

• This toy 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

• In the full package only a stratified sample was executed by us; the rest carry the source authors' evidence (upstream_reported). No task is benchmark_eligible yet.
• This toy sample is metadata only (text columns empty) and far too small to train or benchmark a coding agent; it shows the schema and the links between tables.
• Not produced, reviewed or endorsed by the authors of any source dataset.

Citation

SWEGraph v3 (free toy sample). MatrixNorm normalization by williamTLmiller, derived from SWE-bench, SWE-smith, SWE-bench Pro and BigCodeBench. https://huggingface.co/datasets/williamTLmiller/swe-bench-coding-agent-tasks

Contact

energea@energeatech.com