PatchPilot patch-generation dataset

README.md

@@ -0,0 +1,75 @@
+---
+license: other
+task_categories: [text-generation]
+tags: [program-repair, code, swe-bench, sft, patchpilot]
+---
+# PatchPilot patch-generation dataset
+
+Supervised fine-tuning chats for PatchPilot's patch generator (A2). Each chat is exactly the
+prompt PatchPilot's agent sends to its model, followed by the developers' real fix written in the
+agent's SEARCH/REPLACE edit format.
+
+## Source
+
+Built by `scripts/build_patchgen_data.py` (seed 42) from the **SWE-bench training split**
+(`princeton-nlp/SWE-bench`, train) and the gold files' contents from `princeton-nlp/SWE-bench_oracle`.
+The same seeded selection (at most 400 instances per repository) and the same repository-level
+train/validation/test splits as the localization dataset are used.
+
+## Leakage check
+
+The training split's 35 repositories are disjoint from SWE-bench Lite's 12. `scripts/check_leakage.py`
+found zero overlap with all 300 SWE-bench Lite test instances on repository, instance id,
+normalised issue text and normalised gold patch (`results/leakage_check.json`).
+
+## Construction
+
+1. **Filter (same shape as SWE-bench Lite's own selection):** the gold patch edits exactly one
+   existing non-test Python file, in at most three hunks. Larger multi-file fixes are skipped.
+2. **Target:** each hunk becomes a SEARCH/REPLACE block. If the original lines are not unique in
+   the file, real neighbouring lines are added until they are. Applying the blocks is checked to
+   give exactly the same file as `git apply` of the gold patch; examples that fail are skipped.
+3. **Prompt:** the agent's system prompt; a user turn with the issue text (max 3,000 characters),
+   the tests added by the fix (max 1,500 characters) and the gold file with line numbers, whole if
+   at most 150 lines, otherwise windows of 15 lines around each edit.
+4. **Length:** user + assistant at most 10,000 characters (about 2,800 Gemma tokens).
+
+## Size
+
+From `results/patchgen/dataset_stats.json`:
+
+| Split | Chats |
+|---|---|
+| train | 3,063 |
+| val | 494 |
+| test | 580 |
+
+Of the 8,867 selected training-split instances, 4,137 became chats. Skipped: 4,113 not
+SWE-bench-Lite-shaped (more than one file or more than three hunks), 235 too long, 197 with no
+existing non-test Python file edited, 101 absent from `SWE-bench_oracle`, 78 whose diff could not
+be turned into unique SEARCH/REPLACE blocks that reproduce `git apply`, and 6 whose edited file was
+missing from the oracle text. User + assistant length: median 5,207 characters, maximum 9,995.
+The splits use the same repositories as the localization dataset.
+
+## Format
+
+`{train,val,test}.jsonl.gz`, one chat per line:
+
+```json
+{"instance_id": "...", "repo": "owner/name", "split": "train", "files": ["pkg/mod.py"],
+ "messages": [{"role": "system", "content": "..."},
+              {"role": "user", "content": "## Bug report ..."},
+              {"role": "assistant", "content": "Fix:\n\npkg/mod.py\n<<<<<<< SEARCH\n..."}]}
+```
+
+## Limitations
+
+- The context shows the gold file around the edit, so the model learns to fix given good
+  localization; at run time it sees the agent's localized files, which may be wrong.
+- The tests are shown as the test patch, not as the failing test output the agent sees at run time.
+- Only small, single-file fixes; the developers' fix is one correct answer among possibly many.
+
+## Licence
+
+Derived from SWE-bench (MIT). The underlying code belongs to the respective open-source projects
+under their own licences.

dataset_stats.json

@@ -0,0 +1,63 @@
+{
+  "source": [
+    "princeton-nlp/SWE-bench",
+    "princeton-nlp/SWE-bench_oracle"
+  ],
+  "seed": 42,
+  "per_repo_cap": 400,
+  "selected_instances": 8867,
+  "examples": {
+    "train": 3063,
+    "test": 580,
+    "val": 494
+  },
+  "skipped": {
+    "not_lite_like": 4113,
+    "no_oracle": 101,
+    "too_long": 235,
+    "no_python_source_edit": 197,
+    "not_convertible": 78,
+    "missing_source": 6
+  },
+  "user_plus_assistant_chars": {
+    "median": 5207,
+    "max": 9995
+  },
+  "repo_split": {
+    "conda/conda": "test",
+    "explosion/spaCy": "test",
+    "apache/airflow": "test",
+    "mesonbuild/meson": "test",
+    "tiangolo/fastapi": "val",
+    "open-mmlab/mmdetection": "val",
+    "scipy/scipy": "val",
+    "huggingface/transformers": "val",
+    "pypa/pip": "val",
+    "pantsbuild/pants": "train",
+    "google/jax": "train",
+    "dagster-io/dagster": "train",
+    "docker/compose": "train",
+    "python/typeshed": "train",
+    "conan-io/conan": "train",
+    "kubeflow/pipelines": "train",
+    "tensorflow/models": "train",
+    "apache/mxnet": "train",
+    "pyca/cryptography": "train",
+    "DataDog/integrations-core": "train",
+    "wagtail/wagtail": "train",
+    "gitpython-developers/GitPython": "train",
+    "jupyterlab/jupyterlab": "train",
+    "Lightning-AI/lightning": "train",
+    "twisted/twisted": "train",
+    "ray-project/ray": "train",
+    "numpy/numpy": "train",
+    "PrefectHQ/prefect": "train",
+    "pandas-dev/pandas": "train",
+    "Qiskit/qiskit": "train",
+    "ytdl-org/youtube-dl": "train",
+    "googleapis/google-cloud-python": "train",
+    "ipython/ipython": "train",
+    "JohnSnowLabs/spark-nlp": "train",
+    "celery/celery": "train"
+  }
+}

leakage_check.json

@@ -0,0 +1,96 @@
+{
+  "train_source": "princeton-nlp/SWE-bench",
+  "train_instances_checked": 8867,
+  "train_repos": [
+    "DataDog/integrations-core",
+    "JohnSnowLabs/spark-nlp",
+    "Lightning-AI/lightning",
+    "PrefectHQ/prefect",
+    "Qiskit/qiskit",
+    "apache/airflow",
+    "apache/mxnet",
+    "celery/celery",
+    "conan-io/conan",
+    "conda/conda",
+    "dagster-io/dagster",
+    "docker/compose",
+    "explosion/spaCy",
+    "gitpython-developers/GitPython",
+    "google/jax",
+    "googleapis/google-cloud-python",
+    "huggingface/transformers",
+    "ipython/ipython",
+    "jupyterlab/jupyterlab",
+    "kubeflow/pipelines",
+    "mesonbuild/meson",
+    "numpy/numpy",
+    "open-mmlab/mmdetection",
+    "pandas-dev/pandas",
+    "pantsbuild/pants",
+    "pyca/cryptography",
+    "pypa/pip",
+    "python/typeshed",
+    "ray-project/ray",
+    "scipy/scipy",
+    "tensorflow/models",
+    "tiangolo/fastapi",
+    "twisted/twisted",
+    "wagtail/wagtail",
+    "ytdl-org/youtube-dl"
+  ],
+  "all_train_split_instances_checked": 19008,
+  "eval_source": "princeton-nlp/SWE-bench_Lite",
+  "eval_instances_checked": 300,
+  "eval_repos": [
+    "astropy/astropy",
+    "django/django",
+    "matplotlib/matplotlib",
+    "mwaskom/seaborn",
+    "pallets/flask",
+    "psf/requests",
+    "pydata/xarray",
+    "pylint-dev/pylint",
+    "pytest-dev/pytest",
+    "scikit-learn/scikit-learn",
+    "sphinx-doc/sphinx",
+    "sympy/sympy"
+  ],
+  "eval_subset_instances": 50,
+  "selected_train_vs_eval": {
+    "repo": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "instance_id": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "issue_text": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "gold_patch": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    }
+  },
+  "full_train_split_vs_eval": {
+    "repo": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "instance_id": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "issue_text": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    },
+    "gold_patch": {
+      "overlapping_eval_instances": 0,
+      "examples": []
+    }
+  },
+  "leak_free": true
+}

test.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d328feb0c34c133c0ef6c2095bf783a092e0882f70966abb3d96d29238a160b
+size 824841

train.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47ae221ac41da451f045cee9bb11b2c95bdb9b6474c6ff99c3d5b5de157cccf7
+size 3898484

val.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:187251a5b2eff4b3c7c3c58267b6b36fa34c2ec8354b7cf0d2d129166ce5254b
+size 683832