SWE-bench-Science / README.md
Kevin589981's picture
Publish licensing audit and Docker Hub digests for tasks 077-090
a319eb2 verified
|
Raw History Blame
6.79 kB
metadata
pretty_name: SWE-bench Science
language:
  - en
task_categories:
  - text-generation
tags:
  - software-engineering
  - code-agents
  - scientific-computing
  - benchmark
license: mit

SWE-bench Science

SWE-bench Science evaluates coding agents on software-engineering problems from scientific computing repositories. The release contains 119 tasks. Each task has an isolated environment image and a separate verifier image, both pinned to immutable linux/amd64 Docker Hub digests.

Dataset Summary

Metric Value
Total tasks 119
Default unrestricted-license tasks 100
Restricted-license opt-in tasks 19
GPL/LGPL/AGPL-family tasks 18
Task environment images 119 Docker Hub digests
Separate verifier images 119 Docker Hub digests
linux/amd64 coverage 119/119

The complete generated tables are in data/tasks.csv and data/statistics.md. License values are recorded from the upstream repository, source metadata, or source file headers; non-SPDX family labels are kept explicit when a project does not use a standard SPDX id.

What Is In The Dataset

The snapshot contains the task table, statistics, fixed selections, thin Harbor/Pier task bundles, and the helper tools under tools/. It does not contain reference-answer patches, credentials, agent trajectories, or private verifier tests. The environment image contains the baseline source and public dependencies. The verifier image contains held-out tests and the grader.

Pier installs the selected agent harness at evaluation time. Codex, Claude Code, mini-swe-agent, and other Pier-supported agents are evaluation configuration, not per-task image variants.

Quick Start

python3 -m pip install "huggingface_hub[cli]"
hf auth login
hf download OpenMOSS-Team/SWE-bench-Science \
  --repo-type dataset --local-dir swe-bench-science
cd swe-bench-science

uv tool install "datacurve-pier==0.3.0"
# Without uv: python3 -m pip install "datacurve-pier==0.3.0"
docker login

Materialize the default 100-task selection:

python3 tools/materialize.py --output tasks-selected --force

Materialize a specific list, including inclusive ranges:

python3 tools/materialize.py \
  --task-id 002,005-007 \
  --output tasks-selected-small --force

Restricted-license tasks require an explicit gate. This includes GPL-family tasks and task 019, whose upstream license is academic non-commercial:

python3 tools/materialize.py \
  --allow-restricted-licenses \
  --output tasks-selected-all --force

The 18 GPL-family ids are 003, 020, 021, 023, 032, 057, 066, 074, 075, 082, 083, 084, 085, 096, 097, 098, 100, 118. Task 019 uses an academic non-commercial license. All 19 ids require --allow-restricted-licenses. The materializer records the exact result in selection.json.

Run An Evaluation

Install Pier and run a no-op infrastructure smoke:

pier run -p tasks-selected-small \
  --agent nop --env docker --n-concurrent 1 \
  --no-force-build --no-delete --yes

Run Claude Code, Codex, or mini-swe-agent by changing only the harness, model, and provider profile:

# Claude Code
pier run -p tasks-selected-small --agent claude-code --env docker \
  --env-file ~/.config/swe-bench-science/claude.env \
  --model anthropic/claude-opus-4-7 --n-concurrent 1

# mini-swe-agent
pier run -p tasks-selected-small --agent mini-swe-agent --env docker \
  --env-file ~/.config/swe-bench-science/mini-swe-agent.env \
  --model openai/gpt-5 --n-concurrent 1

The repository also provides an optional wrapper that pre-pulls every immutable image and writes a flat summary:

python3 tools/run_batch.py \
  --path tasks-selected-small \
  --agent codex \
  --env-file ~/.config/swe-bench-science/codex.env \
  --n-concurrent 2 --n-attempts 1 \
  --jobs-dir jobs --job-name codex-small

For a short agent smoke, add --agent-timeout-multiplier 0.0223 (approximately 120 seconds against the default task timeout). This only limits the agent stage; scientific verifier builds and tests retain their own timeout.

Provider Profiles

The examples in profiles/ are templates only. Copy them outside the checkout and set permissions to 600:

mkdir -p ~/.config/swe-bench-science
cp profiles/codex.env.example ~/.config/swe-bench-science/codex.env
cp profiles/claude.env.example ~/.config/swe-bench-science/claude.env
cp profiles/mini-swe-agent.env.example ~/.config/swe-bench-science/mini-swe-agent.env
chmod 600 ~/.config/swe-bench-science/*.env

For Codex, set MODEL, OPENAI_API_KEY, CODEX_BASE_URL, CODEX_WIRE_API=responses|chat, CODEX_VERSION, and CODEX_REASONING_EFFORT. CODEX_WIRE_API=responses selects the OpenAI Responses API; chat selects Chat Completions. For Claude Code, set ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL, and optional ANTHROPIC_CUSTOM_HEADERS. mini-swe-agent uses the provider variables expected by its selected model adapter, commonly OPENAI_API_KEY and OPENAI_BASE_URL.

The credential file is read by Pier at runtime. It is not written to task metadata, Dockerfiles, image layers, or result summaries. For Codex profiles using CODEX_BASE_URL or CODEX_WIRE_API, use tools/run_batch.py: it translates those fields into the Codex config_toml required by the gateway.

Results

Pier writes one aggregate result and one trial directory per task/attempt:

jobs/<job-name>/result.json
jobs/<job-name>/<task>__<trial>/verifier/reward.json
jobs/<job-name>/<task>__<trial>/verifier/ctrf.json
jobs/<job-name>/<task>__<trial>/verifier/test-stdout.txt

The wrapper additionally writes jobs/summary.json and jobs/summary.csv. For direct Pier runs, generate them with:

python3 tools/summarize_results.py --jobs-dir jobs

Use pier view jobs to inspect trajectories. Check result.json, reward.json, and test-stdout.txt together when an agent times out or a candidate fails.

Reproducibility And Licensing

selection.json records the exact task ids. Each row in data/tasks.csv records the upstream repository, base commit, implementation language, source license, material licenses and provenance, environment digest, and verifier digest. The project tools and metadata follow the release repository's MIT terms; task source and fixtures retain their upstream licenses. Dataset-level notices are in NOTICE.md, and affected thin task bundles also include fixtures/PROVENANCE.md.

The dataset is independent of GitHub at runtime. After this snapshot is downloaded, task materialization and evaluation use the local bundle plus Docker Hub image digests; no task definition is fetched from the authoring repository.