XLearning-SCU commited on
Commit
abf485a
·
verified ·
1 Parent(s): 480e825

Add Active-SWE dataset card

Browse files
Files changed (1) hide show
  1. README.md +131 -3
README.md CHANGED
@@ -1,5 +1,133 @@
1
  ---
2
- license: other
3
- license_name: mixed-cc-by-4.0-and-upstream
4
- license_link: LICENSE
 
 
 
 
 
 
 
 
 
5
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ - zh
5
+ pretty_name: Active-SWE
6
+ task_categories:
7
+ - software-development
8
+ tags:
9
+ - code
10
+ - bug-fixing
11
+ - code-review
12
+ - software-engineering
13
+ - docker
14
  ---
15
+
16
+ # Active-SWE
17
+
18
+ Active-SWE is a proactive bug-fixing benchmark for evaluating whether a coding
19
+ model can find and repair functional bugs from a set of files selected for
20
+ review. It is designed for local Docker-based evaluation and is different from
21
+ issue-based benchmarks that provide a complete issue description as the main
22
+ input.
23
+
24
+ The benchmark evaluates two kinds of instances:
25
+
26
+ - **simple**: an instance centered on one source bug.
27
+ - **hard**: an instance that combines multiple source bugs and supports
28
+ multi-bug validation.
29
+
30
+ The `set` column identifies the instance type. The same `instance_id` can occur
31
+ once in each set, so use `(set, instance_id)` as the stable key when combining
32
+ the two sets.
33
+
34
+ ## Files
35
+
36
+ This dataset repository contains Parquet files only:
37
+
38
+ | File | Rows | simple | hard | Use |
39
+ | --- | ---: | ---: | ---: | --- |
40
+ | `Active-SWE.parquet` | 400 | 300 | 100 | Main evaluation set |
41
+ | `Active-SWE-Extend.parquet` | 1663 | 1411 | 252 | Extended evaluation set |
42
+
43
+ The extended set contains the larger evaluation pool. The main set is intended
44
+ for a smaller, faster evaluation run.
45
+
46
+ ## Data fields
47
+
48
+ | Field | Description |
49
+ | --- | --- |
50
+ | `instance_id` | Public instance identifier |
51
+ | `repo` | Source repository in `owner/name` form |
52
+ | `base_commit` | Repository commit used as the evaluation base |
53
+ | `patch` | Ground-truth patch for oracle and evaluation use |
54
+ | `set` | `simple` or `hard` |
55
+ | `language` | Main programming language |
56
+ | `merged_at` | Source change timestamp, when available |
57
+ | `image` | Docker image used by the evaluation harness |
58
+ | `files_pending_review` | Files provided to the proactive review stage |
59
+ | `major_category` | Bug taxonomy labels |
60
+ | `test_framework` | Test framework used by the instance |
61
+ | `log_parser` | Harness log parser for the test output |
62
+ | `template` | Template used to render the evaluation command |
63
+ | `eval_script` | Instance-specific evaluation script |
64
+ | `FAIL_TO_PASS` | Tests expected to pass after the ground-truth fix |
65
+ | `PASS_TO_PASS` | Tests expected to remain passing |
66
+ | `source_f2p` | Source-level F2P mapping, mainly used for hard instances |
67
+ | `source_p2p` | Source-level P2P mapping, mainly used for hard instances |
68
+ | `edit_hunks` | Changed-file and hunk metadata |
69
+
70
+ The Parquet files include oracle and evaluation fields because they are needed
71
+ by the harness to calculate the benchmark metrics. They should not be exposed
72
+ as model input during generation. In particular, `patch`, `eval_script`,
73
+ `FAIL_TO_PASS`, `PASS_TO_PASS`, and source-level oracle fields are evaluation
74
+ inputs, not prompt fields for the proactive review model.
75
+
76
+ ## Images
77
+
78
+ The `image` column contains the exact Docker image reference used by the
79
+ release harness. Images are hosted in the public Docker Hub repository
80
+ `docker.io/biningbin/active-swe` and are tagged with the `simple-` or `hard-`
81
+ prefix. The dataset does not contain image layers; pull the referenced image
82
+ before running an evaluation.
83
+
84
+ For example:
85
+
86
+ ```bash
87
+ docker pull docker.io/biningbin/active-swe:simple-pylint-dev__pylint-5417
88
+ ```
89
+
90
+ ## Loading the dataset
91
+
92
+ ```python
93
+ from datasets import load_dataset
94
+
95
+ dataset = load_dataset(
96
+ "parquet",
97
+ data_files={
98
+ "main": "https://huggingface.co/datasets/XLearning-SCU/Active-SWE/resolve/main/Active-SWE.parquet",
99
+ "extend": "https://huggingface.co/datasets/XLearning-SCU/Active-SWE/resolve/main/Active-SWE-Extend.parquet",
100
+ },
101
+ )
102
+ ```
103
+
104
+ Or download one file with the Hugging Face CLI:
105
+
106
+ ```bash
107
+ hf download XLearning-SCU/Active-SWE Active-SWE.parquet \
108
+ --repo-type dataset --local-dir ./active_swe_data
109
+ ```
110
+
111
+ ## Evaluation protocol
112
+
113
+ The release harness runs the benchmark stages locally with Docker. The model
114
+ first receives the files pending review and generates a code patch. Later
115
+ stages can generate and validate tests, and the harness runs the rendered
116
+ evaluation scripts inside the referenced image.
117
+
118
+ The primary paper-facing metrics are:
119
+
120
+ ```text
121
+ LR | LP | Resolved | Count | TV | Revealed
122
+ ```
123
+
124
+ The denominator is the number of rows supplied to the corresponding run. The
125
+ release harness preserves one output row per input instance, including rows
126
+ that fail during generation or evaluation.
127
+
128
+ ## Citation and source licenses
129
+
130
+ Please cite the Active-SWE paper when using this dataset. The source
131
+ repositories represented in the benchmark retain their respective licenses;
132
+ users are responsible for complying with those licenses when inspecting or
133
+ redistributing source code and patches.