README.md

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

<img src="https://huggingface.co/datasets/AdaMLLab/TurMix/resolve/main/finetasks_turkish_main_results.png" width="900" alt="Finetasks benchmark scores, showing TurMix-Matched as SOTA.">

<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>

TurMix ([https://arxiv.org/abs/2512.18834](https://arxiv.org/abs/2512.18834)) is a Turkish pretraining corpus containing 168 billion tokens across 219 million documents (in the minhash subset). Rather than scraping the web again, TurMix combines five publicly available Turkish datasets, applies Turkish-specific quality filtering, and performs cross-dataset deduplication.

We train a 1.4B parameter language model through nanotron on 30 billion tokens to show that the `matched` subset of TurMix outperforms the previous state-of-the-art, [FineWeb-2 Turkish](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) (see [Appendix A9 in the Fineweb-2 paper](https://arxiv.org/pdf/2506.20920)), achieving a 5.5% relative improvement. Furthermore, the `minhash_deduped` subset performs competitively with over 2× the total number of tokens.

## Subsets

| Subset | Documents | Tokens | Description |
|--------|-----------|--------|-------------|
| `quality_filtered` | 394.0M | 307.2B | Quality-filtered data before deduplication |
| `minhash_deduped` | 219.1M | 167.6B | Document-level MinHash deduplication |
| `matched` | 67.6M | 56.0B | 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/TurMix", "minhash_deduped")
ds = load_dataset("AdaMLLab/TurMix", "quality_filtered")
ds = load_dataset("AdaMLLab/TurMix", "matched")
```

## Sources

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

| Source | Tokens (MinHash) | Documents (MinHash) |
|--------|------------------|---------------------|
| HPLT 2.0 | 46.0B | 53.7M |
| FineWeb-2 | 41.9B | 54.5M |
| CulturaX | 35.8B | 47.9M |
| C4 | 25.3B | 36.5M |
| VNGRS-Web | 18.7B | 26.5M |
| **Total** | **167.6B** | **219.1M** |

## Pipeline

1. Quality filtering with Turkish-specific thresholds (terminal punctuation, repetition patterns, Latin script ratio, 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.