Logics-SWE-Env-2.5K

2,553 software engineering task instances · 1,771 repositories · 4 programming languages

🤗 Related model: Logics-SWE-Qwen3.6-27B

📄 Paper: One to More, More to One

💻 GitHub: AgenticBigBang

Overview

What is this dataset?

Logics-SWE-Env-2.5K is a collection of repository-level software engineering tasks for research on coding agents and environment-based reinforcement learning. It contains 2,553 unique task instances from 1,771 GitHub repositories, spanning Go, Python, TypeScript, and JavaScript.

This release is a public subset of the training data used in One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents. It is released as part of the AgenticBigBang project, alongside research on category-aware expert training and multi-teacher on-policy distillation.

Each instance provides a problem statement, the target repository and base revision, a reference code patch, a test patch, and environment-image metadata. These fields support repository-level task setup and test-based verification when used with the corresponding execution environment and runner. This is task and environment data, not an agent-trajectory or SFT demonstration dataset.

Relationship to the paper

The paper One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents develops category experts through Refresh–Repair–Expand (RRE) and integrates them into a single student using multi-teacher on-policy distillation (MOPD). This dataset exposes a subset of the underlying training tasks; it does not contain the complete training corpus, training trajectories, or all stage-specific sampling and balancing decisions.

The related model is Logics-SWE-Qwen3.6-27B. Results reported for that model concern the paper's full training procedure and should not be interpreted as results from training exclusively on this 2.5K subset.

Data Distribution

Programming languages

Task language	Instances	Share
Go	1,377	53.94%
Python	580	22.72%
TypeScript	465	18.21%
JavaScript	131	5.13%
Total	2,553	100.00%

Task language denotes the primary programming language modified by the task patch, rather than necessarily the primary language of the repository. Assignments use reference-patch file extensions for 2,546 instances, test-patch file extensions for 5 instances, and a verified repository-language fallback for 2 instances. Multi-language repositories may have different repository and task languages.

Repository coverage

Statistic	Value
Unique repositories	1,771
Mean instances per repository	1.44
Median instances per repository	1
Repositories contributing one instance	1,403
Largest repository contribution	24 instances (0.94%)
Top-10 repository share	5.68%
Top-20 repository share	8.81%

The repository distribution is long-tailed: 79.22% of repositories contribute exactly one instance. This release is not balanced across programming languages or repositories.

Machine-readable summaries are available in language_distribution.csv, repository_distribution.csv, and summary.json.

Data Format

The data uses JSON Lines, with one task instance per record. All 2,553 records share the following 16 fields and JSON value types:

Field	Type	Description
instance_id	string	Unique task identifier.
repo_name	string	Source repository in owner/repository form.
base_commit	string	Repository revision used as the task's starting point.
problem_statement	string	Task description presented to the agent.
patch	string	Reference code changes; not an agent input.
test_patch	string	Test changes associated with task verification.
docker_image	string	Task-environment container image reference hosted on GHCR (ghcr.io).
image_info	string	Additional environment-image metadata.
modified_files	string	Stored description of modified files.
modified_entity_summaries	string	Stored summaries of modified code entities.
num_non_test_files	number	Non-test file-change count.
num_non_test_func_methods	number	Non-test function/method-change count.
num_non_test_lines	number	Non-test line-change count.
passrate	string	Historical pass-rate metadata; not a model-independent difficulty score.
created_at	string	Source timestamp metadata.
idx	number	Source record index.

The language statistics above are derived annotations; language is not a field in the source records. The source snapshot also does not provide complete per-instance A/B/C expert-category labels.

Quick Start

Load from Hugging Face

Install the Hugging Face datasets package, then load the dataset:

from datasets import load_dataset

tasks = load_dataset(
    "Logics-MLLM/Logics-SWE-Env-2.5K",
    split="train",
)

print(f"Number of tasks: {len(tasks)}")
task = tasks[0]
print(task["instance_id"])
print(task["repo_name"])
print(task["problem_statement"])

The train split contains all 2,553 released instances in data/train.jsonl; no held-out evaluation split is provided. Loading the records does not itself provision task environments or execute verification.

To load a downloaded copy instead, use load_dataset("json", data_files="train.jsonl", split="train").

Use with an agent runner

1. Resolve the task's environment image and initialize the repository at base_commit.
2. Give the agent the problem statement and the allowed repository context.
3. Collect the agent's proposed code changes.
4. Verify those changes using the task-specific tests and runner protocol in an isolated environment.

Keep reference patches, hidden test content, and derived solution summaries out of agent-visible inputs when measuring task-solving performance. Do not evaluate an agent in a workspace where the reference solution has already been applied.

Scope and Limitations

• Training subset, not a held-out benchmark. This release is drawn from the paper's training data. Performance on these tasks is not evidence of held-out generalization for models trained on them.
• Environment availability matters. Image metadata alone does not guarantee that an image is publicly retrievable or compatible with a particular runner. Use verified release image references and check runtime requirements before launching jobs.
• Historical pass rates are protocol-dependent. They depend on the model, sampling, and verification setup used to collect them, and should not be treated as scores of the released model.
• Coverage is uneven. Go accounts for over half of this subset; conclusions should account for its language and repository composition.
• Execute in isolation. Repository code and tests should run in appropriately isolated environments without access to host credentials or unrelated data.

License

The database is made available under the Open Data Commons Attribution License (ODC-BY) v1.0. Please attribute Logics-SWE-Env-2.5K and retain the applicable license notices when using or redistributing the database in accordance with its terms.

This license applies to rights in the database, not to all rights in its individual contents. Repository-derived code, reference patches, test patches, and other third-party materials remain subject to their applicable upstream licenses and notices. Container images and their dependencies may also carry separate licenses. ODC-BY does not replace those licenses or grant additional rights in third-party content.

Citation

If you use this dataset, please cite the associated work:

@misc{zhao2026onetomoremoretoone,
  title         = {One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents},
  author        = {Jie Zhao and Ziyu Jiang and Suhang Zheng and Minghui Shan and Xiaoxiao Xu and Lin Qu},
  year          = {2026},
  eprint        = {2609.23377},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2609.23377}
}

Acknowledgements

We thank the authors and maintainers of the upstream repositories for their open-source contributions.