---
license: mit
task_categories:
- text-generation
language:
- zh
- en
tags:
- ARP
- RL
- verl
- LLM
pretty_name: DeCoMo
size_categories:
- 10K<n<100K
---

---

## 📁 Dataset Variants
### 1. `preprocess-CodeContests-perturbation-none`

* No perturbation applied
* Pure clean programs
* Used as baseline reference

### 2. `preprocess-CodeContests-perturbation-syntax`
* Only syntax-level perturbations
* Bug types: token deletion, indentation errors, malformed expressions
* Focus: **syntax recovery ability**

### 3. `preprocess-CodeContests-perturbation-structure`

* Only structure-level perturbations
* Bug types: line swap, line deletion, statement reordering
* Focus: **program structure repair**

### 4. `preprocess-CodeContests-perturbation-semantic`

* Only semantic-level perturbations
* Bug types: logic errors, wrong operators, boundary mistakes
* Focus: **semantic reasoning repair**

### 5. `preprocess-CodeContests-perturbation-mix`

* Mixed perturbation dataset
* Combines syntax + structure + semantic corruption
* Used for general robustness training

### 6. `preprocess-CodeContests-perturbation-mixed-task-*`

These are **controlled mixture ratio subsets** used for ablation and scaling.

Format:

```
00100 → 0 : 10 : 0 : 0
01000 → 0 : 0 : 10 : 0
10000 → 10 : 0 : 0 : 0
2242  → 2 : 2 : 4 : 2
```

Where the order is:

```
[syntax : structure : semantic : mixed]
```

Examples:

* `00100` → only structure perturbation (100%)
* `01000` → only semantic perturbation (100%)
* `10000` → only syntax perturbation (100%)
* `2242` → balanced mixture (2:2:4:2)

---

## 📦 Data Format

Each sample contains:

```json id="m4kqv2"
{
  "id": "problem_id",
  "clean_code": "correct reference solution",
  "buggy_code": "perturbed program",
  "perturbation_type": "syntax | structure | semantic | mixed",
  "language": "python",
  "task_prompt": "problem statement (optional)",
  "test_cases": {
    "input": "...",
    "output": "..."
  }
}
```

## 🎯 Purpose
These variants are designed for:
* perturbation-driven program repair
* RL-based trajectory learning
* ablation over bug types
* robustness evaluation under different corruption distributions