README.md

---
license: mit
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: eval
    path: data/eval-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: repo
    dtype: string
  - name: fix_commit
    dtype: string
  - name: buggy_commit
    dtype: string
  - name: message
    dtype: string
  - name: files
    list:
    - name: path
      dtype: string
    - name: patch
      dtype: string
    - name: additions
      dtype: int64
    - name: deletions
      dtype: int64
    - name: language
      dtype: string
  - name: timestamp
    dtype: timestamp[s]
  splits:
  - name: train
    num_bytes: 1561640639
    num_examples: 115096
  - name: eval
    num_bytes: 29054081
    num_examples: 3000
  - name: test
    num_bytes: 29054081
    num_examples: 3000
  download_size: 549629363
  dataset_size: 1619748801
task_categories:
- text-generation
- summarization
language:
- en
tags:
- code
pretty_name: Github issues dataset
size_categories:
- 100K<n<1M
---

# GitHub Pull Request Bug–Fix Dataset

[Kaggle url](https://www.kaggle.com/datasets/adhavanyuvaraj/github-issues)

A **curated, high-signal dataset of real-world software bugs and fixes** collected from **25 popular open-source GitHub repositories**.  
Each entry corresponds to a **single pull request (PR)** and pairs contextual metadata with the **exact code changes (unified diffs)** that fixed the bug.

This dataset is designed for:

- **Automated program repair**
- **Bug-fix patch generation**
- **LLM-based code and debugging agents**
- **Empirical software engineering research**

---

## How to use

install datasets python library:

```bash
pip install datasets
```
here is a copy paste example
```python
from datasets import load_dataset

# Load all splits
dataset = load_dataset("helloadhavan/github_issues")

print(dataset)
# pick the train split

example = dataset["train"][0]

# Inspect a single example

print("Repository:", example["repo"])
print("Buggy commit:", example["buggy_commit"])
print("Fix commit:", example["fix_commit"])
print("Message:", example["message"])
print("Timestamp:", example["timestamp"])

print("\nModified files:")
for f in example["files"]:
    print("-", f["path"], f["language"])

# Filter examples by programming language

def contains_assembly_file(example):
    return any(f["language"] == "Assembly" for f in example["files"])

python_fixes = dataset["train"].filter(contains_assembly_file)

print("Assembly-related fixes:", len(python_fixes))

```


## Data collection methodology

Data was collected from **GitHub repositories** by identifying commit pairs that represent
a **bug-introducing version** and its corresponding **fix commit**.

The dataset was constructed and post-processed to ensure high signal and usability:

- Only commits representing **bug fixes or correctness changes** were included
- Each example explicitly links a **buggy commit** to the corresponding **fix commit**
- Repository metadata is preserved for traceability
- Code changes are stored as **unified diffs at the file level**
- Commits that only perform refactoring, formatting, or non-functional changes were excluded
- Entries without meaningful code changes were filtered out

Each dataset row represents **one bug–fix commit pair**, rather than a pull request.

---

## Dataset schema

Each entry in the dataset follows the schema below:

```json
{
  "repo": "owner/repository",
  "buggy_commit": "abcdef123456...",
  "fix_commit": "fedcba654321...",
  "message": "Commit message describing the fix",
  "timestamp": "YYYY-MM-DDTHH:MM:SSZ",
  "files": [
    {
      "path": "path/to/file.ext",
      "patch": "unified diff representing the fix",
      "additions": 10,
      "deletions": 2,
      "language": "Programming language inferred from file extension"
    }
  ]
}
```

| Field               | Description                                           |
| ------------------- | ----------------------------------------------------- |
| `repo`              | GitHub repository containing the fix                  |
| `buggy_commit`      | Commit introducing or containing the bug              |
| `fix_commit`        | Commit that fixes the bug                             |
| `message`           | Commit message associated with the fix                |
| `timestamp`         | Timestamp of the fix commit (ISO 8601 format)         |
| `files`             | List of files modified by the fix                     |
| `files[].path`      | Path to the modified file                             |
| `files[].patch`     | Unified diff containing the code changes              |
| `files[].additions` | Number of lines added                                 |
| `files[].deletions` | Number of lines removed                               |
| `files[].language`  | Programming language inferred from the file extension |


## Supported languages

The dataset contains fixes across multiple programming languages, including (but not limited to):

* JavaScript / TypeScript
* C / C++
* Python
* Rust
* Go
* Java
* Objective-C / Objective-C++ (rare)
* Assembly (very rare. only 638 samples)

Language distribution varies by repository.

## Intended use cases

This dataset is well-suited for:

* Training models to generate patches from real pull request context
* Studying bug-fix patterns across large codebases
* Building autonomous debugging or repair agents
* Research in program repair, code synthesis, and software maintenance

It is not intended for:

* Pull request classification or triage
* Sentiment analysis

## Limitations

The dataset reflects real-world noise from GitHub pull requests
Buggy commit identification is heuristic and may be imperfect
Some fixes involve refactoring or design changes rather than minimal patches
No guarantee that fixes represent optimal or best-practice solutions

<blockquote
  style="
    background: #fff7cc;
    border-left: 5px solid #ffad00;
    padding: 12px 16px;
    color: #5c4b00;
    font-style: italic;
    border-radius: 4px;
  "
>
  <strong style="color:rgba(57, 0, 0, 1)">Note:</strong> Due to a bug in the scraper code, 121k samples were collected instead of the planned 50k.
</blockquote>