README.md

---
license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - code
  - software-engineering
  - swe-bench
---

# SWE-rebench-V2 clean Python tasks

A train/test split of Python tasks from [nebius/SWE-rebench-V2](https://huggingface.co/datasets/nebius/SWE-rebench-V2).
We took the Python subset of the original dataset and kept only the tasks where the golden patch passes
the unit tests and the empty patch does not.

- **train**: 3,837 instances from 408 repositories
- **test**: 500 instances from 100 repositories

The split is made by repository, so no repository appears in both splits.

![Train repo distribution](train_repo_distribution.png)

![Test repo distribution](test_repo_distribution.png)

We evaluated multiple models on the test split as of June 2026 — the comparison is
[here](https://huggingface.co/whitecircle/GLM-4.7-Flash-Coder).