AadiBhatia/code-edit-quality

---
configs:
- config_name: clean
data_files:
- split: train
path: clean/train.parquet
- config_name: dirty
data_files:
- split: train
path: dirty/train.parquet
license: apache-2.0
task_categories:
- text-generation
tags:
- code-editing
- quality-filtering
- sft
- sharegpt
size_categories:
- 10K<n<100K
---

# Code Editing Quality — SFT-Ready (ShareGPT Format)

Quality-filtered splits of a 50K code-editing SFT dataset in **ShareGPT conversation format**, produced by LLM-based distillation that evaluates 9 quality criteria per sample.

## Format

Each sample has a `conversations` field with ShareGPT-style turns:
- **system**: Code editing system prompt
- **human**: Instruction + source code
- **gpt**: Edited code

Compatible with [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl), [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory), and other SFT frameworks that support ShareGPT format.

## Splits

| Split | Samples | Description |
|---|---|---|
| `clean` | 21,774 | Samples with **zero** antipatterns across all 9 criteria |
| `dirty` | 27,773 | Samples with **at least one** antipattern detected |

## Usage

```python
from datasets import load_dataset

clean = load_dataset("AadiBhatia/code-edit-quality", "clean", split="train")
dirty = load_dataset("AadiBhatia/code-edit-quality", "dirty", split="train")

# Each sample:
# clean[0]["conversations"] -> [{system}, {human}, {gpt}]
```