AGENTDARS/reviewer-train

---
license: cc-by-4.0
---
## DATASET SUMMARY
This dataset contains 164k samples of multiple combination of patches to various software engineering repositories along with critiques for each patch. The patches are collected from the repositories of [Nebius's Agent](https://huggingface.co/datasets/nebius/SWE-agent-trajectories) while the critiques are generated using GPT-4o by providing the information about failing test cases.

## How to use
```python
from datasets import load_dataset
dataset = load_dataset("AGENTDARS/reviewer-train")
```

## Dataset Statistics
<table border="1">
<tr>
<th>Statistic</th>
<th>Value</th>
</tr>
<tr>
<td>Total Trajectories</td>
<td>80K</td>
</tr>
<tr>
<td>Unique Issues</td>
<td>3K</td>
</tr>
<tr>
<td>Open-Source Repositories</td>
<td>1,077</td>
</tr>
<tr>
<td>Unique Patches (Total)</td>
<td>42K</td>
</tr>
<tr>
<td>- Positive Patches</td>
<td>7.3K</td>
</tr>
<tr>
<td>- Negative Patches</td>
<td>34.7K</td>
</tr>
<tr>
<td>Unique Issues Correctly Solved</td>
<td>837</td>
</tr>
<tr>
<td>Training Examples Generated</td>
<td>164K</td>
</tr>
<tr>
<td>Total Tokens in Training Data</td>
<td>~500M</td>
</tr>
</table>

## Dataset Fields

The dataset consists of several fields containing structured information about software patches, problem statements, user prompts, solution prompts, and chat data.

### Fields Description

| **Field Name** | **Description** |
|--------------------|----------------|
| `patches` | Contains various candidate patches being compared for a given issue. |
| `problem_statement` | Describes the issue being addressed in the patch, obtained from the GitHub issue page. |
| `instance_id` | A unique identifier assigned to each data instance. |
| `user_prompt` | The instruction given to an AI model or evaluator, describing the task. |
| `solution_prompt` | Critiques generated by the AI model for the patches. |
| `chat_data` | Contains multi-turn dialogues where the model evaluates and critiques patches. |

### License
The dataset is licensed under the Creative Commons Attribution 4.0 license. However, please respect the license of each specific repository on which a particular instance is based.