randomcc commited on
Commit
d26b4b4
·
verified ·
1 Parent(s): acd0c48

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +87 -1
README.md CHANGED
@@ -13,4 +13,90 @@ tags:
13
  pretty_name: DeCoMo
14
  size_categories:
15
  - 10K<n<100K
16
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  pretty_name: DeCoMo
14
  size_categories:
15
  - 10K<n<100K
16
+ ---
17
+
18
+ ---
19
+
20
+ ## 📁 Dataset Variants
21
+ ### 1. `preprocess-CodeContests-perturbation-none`
22
+
23
+ * No perturbation applied
24
+ * Pure clean programs
25
+ * Used as baseline reference
26
+
27
+ ### 2. `preprocess-CodeContests-perturbation-syntax`
28
+ * Only syntax-level perturbations
29
+ * Bug types: token deletion, indentation errors, malformed expressions
30
+ * Focus: **syntax recovery ability**
31
+
32
+ ### 3. `preprocess-CodeContests-perturbation-structure`
33
+
34
+ * Only structure-level perturbations
35
+ * Bug types: line swap, line deletion, statement reordering
36
+ * Focus: **program structure repair**
37
+
38
+ ### 4. `preprocess-CodeContests-perturbation-semantic`
39
+
40
+ * Only semantic-level perturbations
41
+ * Bug types: logic errors, wrong operators, boundary mistakes
42
+ * Focus: **semantic reasoning repair**
43
+
44
+ ### 5. `preprocess-CodeContests-perturbation-mix`
45
+
46
+ * Mixed perturbation dataset
47
+ * Combines syntax + structure + semantic corruption
48
+ * Used for general robustness training
49
+
50
+ ### 6. `preprocess-CodeContests-perturbation-mixed-task-*`
51
+
52
+ These are **controlled mixture ratio subsets** used for ablation and scaling.
53
+
54
+ Format:
55
+
56
+ ```
57
+ 00100 → 0 : 10 : 0 : 0
58
+ 01000 → 0 : 0 : 10 : 0
59
+ 10000 → 10 : 0 : 0 : 0
60
+ 2242 → 2 : 2 : 4 : 2
61
+ ```
62
+
63
+ Where the order is:
64
+
65
+ ```
66
+ [syntax : structure : semantic : mixed]
67
+ ```
68
+
69
+ Examples:
70
+
71
+ * `00100` → only structure perturbation (100%)
72
+ * `01000` → only semantic perturbation (100%)
73
+ * `10000` → only syntax perturbation (100%)
74
+ * `2242` → balanced mixture (2:2:4:2)
75
+
76
+ ---
77
+
78
+ ## 📦 Data Format
79
+
80
+ Each sample contains:
81
+
82
+ ```json id="m4kqv2"
83
+ {
84
+ "id": "problem_id",
85
+ "clean_code": "correct reference solution",
86
+ "buggy_code": "perturbed program",
87
+ "perturbation_type": "syntax | structure | semantic | mixed",
88
+ "language": "python",
89
+ "task_prompt": "problem statement (optional)",
90
+ "test_cases": {
91
+ "input": "...",
92
+ "output": "..."
93
+ }
94
+ }
95
+ ```
96
+
97
+ ## 🎯 Purpose
98
+ These variants are designed for:
99
+ * perturbation-driven program repair
100
+ * RL-based trajectory learning
101
+ * ablation over bug types
102
+ * robustness evaluation under different corruption distributions