---
license: apache-2.0
language:
- en
tags:
- code
- programming
- competitive-programming
- mathematics
- reasoning
- chain-of-thought
- tool-use
size_categories:
- 100K<n<1M
task_categories:
- text-generation
---

# uCoder Dataset

A high-quality, deduplicated dataset for training coding and mathematics language models.

## Dataset Statistics

- **Total Samples**: 420,686
- **Format**: ChatML (messages array)
- **Languages**: Python, JavaScript, C++, Java, and more

## Sources

This dataset merges and cleans data from:

| Source | Samples | Description |
|--------|---------|-------------|
| ByteDance-Seed/Code-Contests-Plus | 10,293 | Competitive programming |
| open-r1/codeforces | 47,558 | Codeforces problems |
| sahil2801/CodeAlpaca-20k | 15,784 | General coding |
| nampdn-ai/tiny-codes | 159,992 | Educational examples |
| stallone/glaive-code-assistant-v3 | 94,732 | Code assistance |
| AlicanKiraz0/Agentic-CoT-Coding-SFT | 429 | Chain-of-thought |
| deepmind/code_contests | 12,943 | AlphaCode problems |

## Features

- Samples with test cases: 23,236
- Samples with reasoning: 18,401


## Processing Pipeline

1. **Raw merge**: 441,838 samples
2. **Exact deduplication**: 431,574 samples
3. **MinHash LSH deduplication**: 425,128 samples
4. **Quality filtering**: 420,686 samples

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("uaytug/ucoder-dataset")

# Each sample has:
# - messages: List of {role, content} for ChatML format
# - language: Programming language
# - difficulty: easy/medium/hard/expert
# - source: Original dataset
# - has_tests: Whether test cases are available
# - test_cases: JSON string of test cases
# - has_reasoning: Whether chain-of-thought is included
# - quality_score: Computed quality score
```

## License

Apache 2.0 (inherits from source datasets)

## Citation

```bibtex
@misc{ucoder-dataset,
  title={uCoder Dataset},
  author={Umut Aytuğ Semerci},
  year={2025},
  url={https://huggingface.co/datasets/uaytug/ucoder-dataset}
}
```