daVinci-Env: Open SWE Environment Synthesis at Scale
Paper • 2603.13023 • Published • 30
Datasets:
instance_id stringlengths 22 94 | responses_create_params dict | agent_ref dict |
|---|---|---|
openswe_oss::Zac-HD__shed-91 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::securefederatedai__openfederatedlearning-821 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::venth__aws-adfs-209 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::frequenz-floss__frequenz-sdk-python-335 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::fugue-project__triad-89 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DistributedProofreaders__guiguts-py-937 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::trinodb__trino-python-client-109 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::bodo-ai__PyDough-317 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::stxnext__deep-next-121 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DigitalSlideArchive__digital_slide_archive-192 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::HDCodePractice__EnglishHelper-58 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::sassoftware__viya4-ark-82 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::huggingface__datasets-7353 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::sparkgeo__stac-fastapi-indexed-141 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::SMPTE__ris-osvp-metadata-camdkit-21 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::espdev__csaps-31 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pyccel__pyccel-841 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ApeWorX__silverback-241 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::descope__python-sdk-491 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::betcode-org__flumine-707 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::dbt-labs__dbt-common-112 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::SolarArbiter__solarforecastarbiter-core-257 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pySmartThings__pysmartthings-195 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::eecs-autograder__autograder-server-503 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::silx-kit__silx-4390 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::getmoto__moto-6828 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::blockscout__mcp-server-223 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::open-contracting__kingfisher-collect-381 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::jlowin__fastmcp-336 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::janluke__cloup-173 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::CrossGL__crosstl-204 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::openforis__sepal_ui-488 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::privacyidea__privacyidea-2824 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::themartorana__python-postmark-87 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::fronzbot__blinkpy-546 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ipython__ipython-14029 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::MasoniteFramework__orm-575 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::neural-bandits__calvera-9 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::leonardt__hwtypes-29 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::cantools__cantools-379 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::godaddy__tartufo-337 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::posit-dev__py-shiny-1388 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::meltano__meltano-8184 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::SWE-agent__SWE-agent-732 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::CUQI-DTU__CUQIpy-602 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::okfn__opendataeditor-1053 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::gdsfactory__gdsfactory-2415 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::oceanprotocol__pdr-backend-1136 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::jsbroks__imantics-53 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::swapper-org__NodeChain-73 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::dwavesystems__dimod-1225 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::amadeus4dev__amadeus-python-176 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::omry__omegaconf-295 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::equinor__ecalc-475 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::elastic__apm-agent-python-1914 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::qLDPCOrg__qLDPC-174 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ACCESS-NRI__access-spack-packages-274 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::huggingface__huggingface_hub-2305 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pylint-dev__astroid-2117 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::microsoft__AzureTRE-2873 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::quantumlib__Cirq-3255 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DataRecce__recce-96 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::aio-libs__janus-102 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::neurostuff__PyMARE-111 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::newrelic__newrelic-python-agent-216 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::scikit-learn__scikit-learn-30616 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pyiron__semantikon-269 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::kellerza__pysma-106 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::conan-io__conan-18485 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::acorg__dark-matter-105 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::log2timeline__dfvfs-716 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DeltaRCM__pyDeltaRCM-202 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::laminlabs__lamindb-3224 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::quantumlib__Qualtran-486 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::SMPTE__ris-osvp-metadata-camdkit-119 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::btschwertfeger__python-kraken-sdk-34 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::cgq-qgc__HydroSensorReader-60 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::opendatacube__datacube-core-1253 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ajstensland__slyther-26 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::opsdroid__opsdroid-1781 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::choderalab__openmmtools-779 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::iai-group__MovieBot-117 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::hail-is__hail-14895 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::TOMToolkit__tom_base-1279 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DerOetzi__solaredge2mqtt-198 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::VirtusLab__git-machete-1060 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pypa__pip-7747 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::jax-ml__jax-31885 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pdfminer__pdfminer.six-727 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::blockscout__mcp-server-188 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ewels__rich-click-113 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::ddps-lab__lambda-optimize-serving-54 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::DiamondLightSource__tickit-180 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::openstenoproject__plover-1364 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::elementsinteractive__twyn-284 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::tektronix__tm_devices-134 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::tconbeer__harlequin-104 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::pypa__pip-12795 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::chainer__chainer-1278 | {
"input": []
} | {
"name": "harbor_agent"
} |
openswe_oss::IBM__javacore-analyser-148 | {
"input": []
} | {
"name": "harbor_agent"
} |
End of preview. Expand in Data Studio
OpenSWE-Harbor — NeMo Gym ready
⚠️ Read this before training: there are TWO sets in here — full (45,316) and filtered (8,875).
Set Tasks Where it is When to use Full 45,316 routing/openswe_oss.jsonl,routing/openswe_other.jsonl, all oftasks/Eval-only, dataset analysis, sweeps where you don't care about RL signal quality Filtered (RL default) 8,875 routing/openswe_oss_filtered.jsonl,routing/openswe_other_filtered.jsonl,filtered_ids.txtUse this for RL training and SFT trajectory collection. How to identify a filtered task at the task level:
task.tomlhasmetadata.in_filtered_set = true(orfalse)- Membership is also mirrored at the top of the repo as
filtered_ids.txt(one id per line, all 8,875)- The
routing/<config>_filtered.jsonlfiles only list filtered tasksWhy filtered = RL default? The upstream
GAIR/OpenSWEships 45,320 tasks raw. GAIR then publishesfiltered_ids.csv(8,876 ids) — the curated subset that survives their difficulty-aware filter (drops unsolvable instances + trivially easy ones) and trajectory-sampling curation (a $576K probe where they sampled real agent rollouts and kept only environments where some teacher policy hit reward > 0). This is the set the OpenSWE paper trains OpenSWE-32B (62.4% SWE-Bench Verified) and OpenSWE-72B (66.0%) on.Why you usually don't want the unfiltered full set for RL:
- Impossible tasks (broken envs, ill-posed issues, mis-specified verifiers) where reward stays at 0 forever → wastes compute, shrinks advantage estimates, biases gradients toward "everything is impossible."
- Trivially easy tasks that saturate at reward 1 immediately → no learning signal.
The filter removes both classes. We still ship the full set so you can run eval sweeps or experiment with a wider pool if you have a reason.
What's excluded entirely: 4 of the upstream 45,320 rows have a broken Dockerfile template (unsubstituted
${PYTHON_VERSION}). 1 of those was in GAIR's filtered ids. We drop those 4. Final counts: 45,316 full, 8,875 filtered.Image registry: the
FROMlines inenvironment/Dockerfileare pre-rewritten to point atdocker.io/rvk7895/openswe-python-{3.7,3.9,3.10,3.11,3.12,3.13}:latest(Docker Hub) for the 6 published base versions, covering 36,186 tasks out of the box. Tasks needing an unpublished base (e.g.3.8, 5,318 tasks) keep the original localFROM openswe-python-X.Yline and are flaggedmetadata.base_image_published = false. See Images section for filtering instructions.
Harbor task tree + NeMo Gym routing JSONL derived from GAIR/OpenSWE. Drop-in for running rollouts with the harbor_agent server in NeMo-Gym.
License & attribution: derivative of GAIR/OpenSWE (AGPL-3.0). All credit to Fu et al. — please cite the original paper if you use this. See Citation at the end.
Contents
.
├── README.md
├── convert_openswe.py # generator script (reproduces this layout from GAIR/OpenSWE)
├── extract.sh # extracts the two archives in-place
├── filtered_ids.txt # 8,876 ids; mirror of GAIR's filtered_ids.csv
├── archives/
│ ├── tasks_openswe_oss.tar.zst # 36,881 task dirs, zstd-compressed
│ └── tasks_openswe_other.tar.zst # 8,435 task dirs, zstd-compressed
└── routing/
├── openswe_oss.jsonl # 36,881 routing rows (full)
├── openswe_other.jsonl # 8,435 routing rows (full)
├── openswe_oss_filtered.jsonl # 7,150 routing rows (RL default)
├── openswe_other_filtered.jsonl # 1,725 routing rows (RL default)
├── openswe_oss_skipped.jsonl # 3 rows excluded (broken-template Dockerfile)
└── openswe_other_skipped.jsonl # 1 row excluded (broken-template Dockerfile)
After running ./extract.sh, you also get:
tasks/
├── openswe_oss/<instance_id>/ # 36,881 directories
│ ├── instruction.md # problem_statement
│ ├── task.toml # Harbor TaskConfig + metadata.in_filtered_set
│ ├── environment/Dockerfile # row.Dockerfile (verbatim)
│ ├── solution/solve.sh # applies row.patch
│ └── tests/
│ ├── test.sh # Harbor verifier wrapper (emits reward.txt)
│ └── openswe_eval.sh # row.eval_script (verbatim)
└── openswe_other/<instance_id>/ # 8,435 directories; solve.sh is a no-op (no gold patch)
Tarballs (rather than 272k loose files) because HuggingFace soft-limits repos to 100k files / 10k per folder. Compressed size is ~500–700 MB total; expanded ~3.1 GB.
Quick start — running rollouts with Harbor
Download this repo:
hf download ritvik-sarvam/openswe-harbor --repo-type dataset --local-dir ./openswe-harborExtract the task tree (one-time, ~30s):
cd openswe-harbor && ./extract.sh(Requires
zstd— install viabrew install zstdorapt-get install zstd.)Add to your
harbor_agent.yaml:harbor_agent: responses_api_agents: harbor_agent: harbor_datasets: openswe_oss: local_dataset_path: "/path/to/openswe-harbor/tasks/openswe_oss" workdir: "/testbed" openswe_other: local_dataset_path: "/path/to/openswe-harbor/tasks/openswe_other" workdir: "/testbed" harbor_agent_name: "claude-code" # or your custom agent harbor_environment_type: "modal" # or "docker", "daytona" harbor_environment_kwargs: {}Collect rollouts — for RL, point at the filtered routing files:
cat routing/openswe_oss_filtered.jsonl routing/openswe_other_filtered.jsonl > routing/train.jsonl ng_collect_rollouts \ +agent_name=harbor_agent \ +input_jsonl_fpath=/path/to/openswe-harbor/routing/train.jsonl \ +output_jsonl_fpath=./rollouts.jsonl \ +num_repeats=4Or use
routing/openswe_oss.jsonlfor the full set.
Routing row schema
{
"instance_id": "openswe_oss::Zac-HD__shed-91",
"responses_create_params": {"input": []},
"agent_ref": {"name": "harbor_agent"}
}
instance_id is <dataset_alias>::<task_name> where the alias matches the OpenSWE config name and the task name is the original instance_id.
Subset choice cheat sheet
| File | Rows | Has gold patch | Has verifier | Oracle smoke-test | Use for |
|---|---|---|---|---|---|
openswe_oss.jsonl |
36,881 | ✓ | ✓ | ✓ | full pool / eval sweeps |
openswe_other.jsonl |
8,435 | ✗ | ✓ | Layer 1 only | full pool, no oracle |
openswe_oss_filtered.jsonl |
7,150 | ✓ | ✓ | ✓ | RL training default |
openswe_other_filtered.jsonl |
1,725 | ✗ | ✓ | Layer 1 only | curated, paper-verified |
Images — Docker Hub bases for 6 versions
Base images for 6 Python versions are published on Docker Hub at docker.io/rvk7895/openswe-python-{3.7,3.9,3.10,3.11,3.12,3.13}:latest. For tasks that use one of these, the environment/Dockerfile's FROM line is already rewritten to point at Docker Hub, so Harbor's Docker/Modal backend can pull the base directly.
Coverage
| Status | Tasks | Bases |
|---|---|---|
| Pullable from Docker Hub (FROM rewritten) | 36,186 | openswe-python-{3.7, 3.9, 3.10, 3.11, 3.12, 3.13} |
| Already pullable (public images) | 616 | continuumio/miniconda3, python:3.X-slim, ubuntu:XX.XX, etc. |
| Need a base NOT yet pushed | 8,514 | openswe-python-3.8 (5,318), 3.6 (1,814), 2.7 (693), 3.5 (476), 3.14 (212), one stray :latest (1) |
| Total functional out-of-the-box | 36,802 | 81% of the dataset |
To filter to just the working subset:
# Routing: keep only tasks whose base is published
python -c "
import json,tomllib
from pathlib import Path
for f in ['routing/openswe_oss.jsonl','routing/openswe_other.jsonl']:
out = open(f.replace('.jsonl','_published.jsonl'),'w')
for line in open(f):
row = json.loads(line)
alias, iid = row['instance_id'].split('::')
cfg = tomllib.loads(Path(f'tasks/{alias}/{iid}/task.toml').read_text())
if cfg['metadata']['base_image_published']:
out.write(line)
"
# → routing/openswe_oss_published.jsonl, routing/openswe_other_published.jsonl
Each task.toml has metadata.base_image_published = true|false for per-task filtering, and metadata.local_image_tag retains the original OpenSWE local tag (e.g. openswe--Zac-HD__shed-91) for reference.
Tasks still need COPY repo /testbed resolved
Even with the base pullable, each task's Dockerfile has COPY repo /testbed, expecting a repo/ subdir at build context. Two ways to satisfy this:
| Option | How | Pros / cons |
|---|---|---|
| GAIR's pipeline | Run scripts/build_images.py from GAIR's OpenSWE repo. Materializes repo/ from cached clones, builds each image, tags it openswe--<id>. |
Faithful to OpenSWE; needs the original SETUP_DIR cache. |
--rewrite-dockerfile |
Re-run this dataset's convert_openswe.py --rewrite-dockerfile --dockerhub-user rvk7895. Swaps COPY repo /testbed for RUN git clone https://github.com/<repo>.git /testbed && git reset --hard <base_commit>. |
Fully self-contained Dockerfiles; needs upstream repo to still be public. |
For the easiest path: re-run convert_openswe.py --rewrite-dockerfile --dockerhub-user rvk7895 after downloading — you get Dockerfiles that need zero local prep, just docker build.
To extend coverage
Push the missing bases following GAIR's prepare_baseimg.py recipe (each builds from continuumio/miniconda3 with a different conda env), then add the version to --dockerhub-versions and re-run the converter. The biggest gap is openswe-python-3.8 covering 5,318 tasks.
Oracle vs. verifier — what works for which subset
Harbor's oracle agent applies solution/solve.sh and checks that the verifier emits reward 1.0. Two layers of sanity check:
- Layer 1: env builds, container starts, verifier runs,
reward.txtgets written. - Layer 2: reward 1.0 is reachable (proves the test suite + reward semantics are correct).
For openswe_other rows, there's no gold patch — solve.sh is a no-op exit 0. You still get Layer 1 (proves env build + plumbing works) but lose Layer 2. The 1,725 openswe_other_filtered.jsonl rows have empirical Layer 2 confirmation from GAIR's trajectory-sampling curation pass (that's what got them into filtered_ids.csv to begin with), so they're safe for RL — they're just not oracle-smoke-testable.
Verifier reward contract
tests/test.sh runs tests/openswe_eval.sh and writes a scalar to /logs/verifier/reward.txt:
1.0if the OpenSWE eval script exits 0 (all tests pass)0.0otherwise
openswe_eval.sh is GAIR's eval_script verbatim. Edit test.sh if you want richer rewards (per-test breakdown, partial credit, etc.) — the raw script is preserved.
Reproducing this dataset
# 1. Request access to GAIR/OpenSWE on the HF page, then:
hf download GAIR/OpenSWE --repo-type dataset --local-dir /tmp/openswe_data
# 2. Convert (full set, both filtered + unfiltered routing files emitted):
python convert_openswe.py \
--source-dir /tmp/openswe_data \
--output-dir .
# 3. Or restrict to filtered only:
python convert_openswe.py --source-dir /tmp/openswe_data --output-dir . --filtered-only
# 4. Or make Dockerfiles self-contained (git clone instead of COPY repo):
python convert_openswe.py --source-dir /tmp/openswe_data --output-dir . --rewrite-dockerfile
Schema notes (worth knowing before training)
These all surfaced from inspecting actual rows (the upstream card lists fields that don't behave the way you'd expect):
FAIL_TO_PASS,PASS_TO_PASS,FAIL_TO_FAIL,PASS_TO_FAILare empty across every row. OpenSWE encodes test logic ineval_script, not as a pre-extracted F2P list. Don't filter or score by them.patchis empty for allopenswe_otherrows (non-redistributable repos).test_patchis empty for ~8% ofopenswe_ossrows and allopenswe_otherrows — but in those cases the test diff is inlined as a heredoc insideeval_script, so the verifier is still self-contained.created_atis'1970-01-01T00:00:00Z'for almost every row — not a real timestamp.license_nameforopenswe_otheris dominated byotherandnone. Be careful if your downstream model artifact is license-restricted.
Citation
Please cite the original OpenSWE paper:
@misc{fu2026davincienvopensweenvironment,
title={daVinci-Env: Open SWE Environment Synthesis at Scale},
author={Dayuan Fu and Shenyu Wu and Yunze Wu and Zerui Peng and Yaxing Huang and Jie Sun and Ji Zeng and Mohan Jiang and Lin Zhang and Yukun Li and Jiarui Hu and Liming Liu and Jinlong Hou and Pengfei Liu},
year={2026},
eprint={2603.13023},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2603.13023}
}
Source: https://huggingface.co/datasets/GAIR/OpenSWE · https://github.com/GAIR-NLP/OpenSWE
- Downloads last month
- 69