README.md

---
language:
- th
license: other
task_categories:
- text-generation
arxiv: 2512.18834
configs:
- config_name: minhash_deduped
  data_files:
  - split: train
    path: minhash_deduped/*.parquet
- config_name: matched
  data_files:
  - split: train
    path: consensus/*.parquet
default: minhash_deduped
---

<p align="center">
  <a href="https://huggingface.co/collections/AdaMLLab/mixminmatch">
    <img src="https://img.shields.io/badge/🤗_Collection-MixMinMatch-blue" alt="MixMinMatch Collection">
  </a>
</p>

ThaiMix ([https://arxiv.org/abs/2512.18834](https://arxiv.org/abs/2512.18834)) is a Thai pretraining corpus containing 70 billion tokens across 81 million documents (in the minhash subset). Rather than scraping the web again, ThaiMix combines five publicly available Thai datasets, applies Thai-specific quality filtering, and performs cross-dataset deduplication.

## Subsets

| Subset | Documents | Tokens | Description |
|--------|-----------|--------|-------------|
| `minhash_deduped` | 81.3M | 70.5B | Document-level MinHash deduplication |
| `matched` | 10.9M | 12.3B | Documents appearing in 2+ source datasets |

The matched subset uses cross-dataset agreement as a signal for quality.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("AdaMLLab/ThaiMix", "minhash_deduped")
ds = load_dataset("AdaMLLab/ThaiMix", "matched")
```

## Sources

Tokens were counted using `meta-llama/Llama-3.2-3B`'s tokenizer.

| Source | Survival Rate |
|--------|---------------|
| HPLT 2.0 | 75.1% |
| C4 | 72.6% |
| SEA-CC | 70.7% |
| CulturaX | 69.0% |
| FineWeb-2 | 61.7% |

## Pipeline

1. Quality filtering with Thai-specific thresholds (Thai script ratio, repetition patterns, language identification)
2. Document-level MinHash deduplication (5-gram shingles, 14 bands, 8 hashes per band, similarity threshold 0.8)
3. Cross-source matching to identify documents appearing in 2+ independent sources

## Citation

```bib
@misc{alrashed2025mixminmatch,
      title={Mix, MinHash, and Match: Cross-Source Agreement for Multilingual Pretraining Datasets}, 
      author={Sultan Alrashed and Francesco Orabona},
      year={2025},
      eprint={2512.18834v2},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.18834v2}, 
}
```

## License

See individual source dataset licenses.