Paper | Project Page | HF Space | Code | Dataset | Docker Images
Active-SWE is a proactive bug-fixing benchmark for evaluating whether a coding model can find and repair functional bugs from a set of files selected for review. It is designed for local Docker-based evaluation and is different from issue-based benchmarks that provide a complete issue description as the main input. The benchmark evaluates two kinds of instances: - **simple**: an instance centered on one source bug. - **hard**: an instance that combines multiple source bugs and supports multi-bug validation. The `set` column identifies the instance type. The same `instance_id` can occur once in each set, so use `(set, instance_id)` as the stable key when combining the two sets. ## Files This dataset repository contains one configuration with two previewable splits: | Split | File | Rows | simple | hard | Use | | --- | --- | ---: | ---: | ---: | --- | | `main` | `data/Active-SWE.parquet` | 400 | 300 | 100 | Main evaluation set | | `extend` | `data/Active-SWE-Extend.parquet` | 1663 | 1411 | 252 | Extended evaluation set | The extended set contains the larger evaluation pool. The main set is intended for a smaller, faster evaluation run. ## Data fields | Field | Description | | --- | --- | | `instance_id` | Public instance identifier | | `repo` | Source repository in `owner/name` form | | `base_commit` | Repository commit used as the evaluation base | | `patch` | Ground-truth patch for oracle and evaluation use | | `set` | `simple` or `hard` | | `language` | Main programming language | | `merged_at` | Source change timestamp, when available | | `image` | Docker image used by the evaluation harness | | `files_pending_review` | Files provided to the proactive review stage | | `major_category` | Bug taxonomy labels | | `test_framework` | Test framework used by the instance | | `log_parser` | Harness log parser for the test output | | `template` | Template used to render the evaluation command | | `eval_script` | Instance-specific evaluation script | | `FAIL_TO_PASS` | Tests expected to pass after the ground-truth fix | | `PASS_TO_PASS` | Tests expected to remain passing | | `source_f2p` | Source-level F2P mapping, mainly used for hard instances | | `source_p2p` | Source-level P2P mapping, mainly used for hard instances | | `edit_hunks` | Changed-file and hunk metadata | The Parquet files include oracle and evaluation fields because they are needed by the harness to calculate the benchmark metrics. They should not be exposed as model input during generation. In particular, `patch`, `eval_script`, `FAIL_TO_PASS`, `PASS_TO_PASS`, and source-level oracle fields are evaluation inputs, not prompt fields for the proactive review model. ## Images The `image` column contains the exact Docker image reference used by the release harness. Images are hosted in the public Docker Hub repository `docker.io/biningbin/active-swe` and are tagged with the `simple-` or `hard-` prefix. The dataset does not contain image layers; pull the referenced image before running an evaluation. For example: ```bash docker pull docker.io/biningbin/active-swe:simple-pylint-dev__pylint-5417 ``` ## Loading the dataset ```python from datasets import load_dataset dataset = load_dataset("XLearning-SCU/Active-SWE", name="main") main_data = dataset["main"] extend_data = dataset["extend"] ``` Or download one file with the Hugging Face CLI: ```bash hf download XLearning-SCU/Active-SWE data/Active-SWE.parquet \ --repo-type dataset --local-dir ./active_swe_data ``` ## Evaluation protocol The release harness runs the benchmark stages locally with Docker. The model first receives the files pending review and generates a code patch. Later stages can generate and validate tests, and the harness runs the rendered evaluation scripts inside the referenced image. The primary paper-facing metrics are: ```text LR | LP | Resolved | Count | TV | Revealed ``` The denominator is the number of rows supplied to the corresponding run. The release harness preserves one output row per input instance, including rows that fail during generation or evaluation. ## Paper **[Active-SWE: Benchmarking Coding Agents for Proactive Bug Fixing without Issue Reports](https://arxiv.org/abs/2608.04682)** Haobin Li, Ping Deng, Weizhong Qian, Liang Jiang, Zhenyu Huang, Mouxing Yang, and Xi Peng. Active-SWE evaluates coding agents on proactively discovering and fixing bugs without issue-report guidance. The benchmark contains 1,663 tasks across six bug categories and eight programming languages, including both single-bug and multi-bug evaluation settings. - Project page: https://hbinli.github.io/Active-SWE/ - Paper: https://arxiv.org/abs/2608.04682 - PDF: https://arxiv.org/pdf/2608.04682 - Code: https://github.com/XLearning-SCU/Active-SWE - Docker images: https://hub.docker.com/r/biningbin/active-swe ## Citation and source licenses Please cite the Active-SWE paper when using this dataset: ```bibtex @article{li2026activeswe, title={Active-SWE: Benchmarking Coding Agents for Proactive Bug Fixing without Issue Reports}, author={Li, Haobin and Deng, Ping and Qian, Weizhong and Jiang, Liang and Huang, Zhenyu and Yang, Mouxing and Peng, Xi}, journal={arXiv preprint arXiv:2608.04682}, year={2026} } ``` The source repositories represented in the benchmark retain their respective licenses; users are responsible for complying with those licenses when inspecting or redistributing source code and patches.