Publish CLB SWE-bench 72-task test split

README.md

@@ -0,0 +1,48 @@
+---
+pretty_name: CLB SWE-bench
+language:
+- en
+task_categories:
+- text-generation
+tags:
+- software-engineering
+- automated-program-repair
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: test.jsonl
+---
+
+# CLB SWE-bench
+
+CLB SWE-bench measures software repair on 72 cross-language bugs. The `test` split contains 50 Python–C cases, 11 Java–C cases with upstream-derived tests, and 11 Java–C cases with tests written by the CLB authors.
+
+## Fields
+
+| Field | Description |
+| --- | --- |
+| `instance_id` | Stable CLB identifier in the form `clb-pyc-NNN` or `clb-javac-NNN`. |
+| `repo` | Upstream repository in `owner/name` format. |
+| `base_commit` | Source revision at which the repair starts. |
+| `problem_statement` | Bug description provided to the repair agent. |
+| `image` | Object with the Docker `pull_reference` pinned by digest and its expected `image_id`. |
+| `cohort` | One of `python50`, `java_upstream11`, and `java_author11`. |
+| `patch` | Reference repair, stored as a Git diff. |
+| `test_patch` | Git diff applied by the evaluator to add or modify tests. |
+| `FAIL_TO_PASS` | Target tests that a successful repair must pass. |
+| `PASS_TO_PASS` | Regression tests that must pass after applying the candidate. |
+
+## Evaluation protocol
+
+For candidate generation, provide the selected `problem_statement` and an isolated checkout at `base_commit`. The evaluator starts from the pinned image, applies the candidate and `test_patch`, then checks both test lists. The `patch` field provides the reference repair for Gold controls.
+
+To load the local data before publishing the dataset, run the following from this directory.
+
+```python
+from datasets import load_dataset
+
+tasks = load_dataset("json", data_files={"test": "test.jsonl"}, split="test")
+```
+
+The release exposes the reference and test patches for reproducibility. Use the generation input projection above when running a benchmark whose agent must work from the problem statement and Base.