README.md

---
language:
- ar
license: other
task_categories:
- text-generation
arxiv: 2512.18834
configs:
- config_name: minhash_deduped
  data_files:
  - split: train
    path: data/minhash_deduped/*
- config_name: matched
  data_files:
  - split: train
    path: data/consensus/*
- config_name: sentence_deduped
  data_files:
  - split: train
    path: data/sentence_deduped/*
default: minhash_deduped
---

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

AraMix ([https://arxiv.org/abs/2512.18834](https://arxiv.org/abs/2512.18834)) is an Arabic pretraining corpus containing 178 billion tokens across 179 million documents (in the minhash subset). Rather than scraping the web again, AraMix combines seven publicly available Arabic datasets, applies Arabic-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 `consensus` subset of AraMix outperforms the previous state-of-the-art, [arabicweb24](https://huggingface.co/datasets/lightonai/ArabicWeb24) (see [Appendix A9 in the Fineweb-2 paper](https://arxiv.org/pdf/2506.20920)) while having more total tokens. Furthermore, the `minhash_deduped` subset performs on-par with nearly 5 times the total number of tokens.

In this setup, we remove all samples in `consensus` with more than 5 duplicates.

## Subsets

| Subset | Documents | Tokens | Description |
|--------|-----------|--------|-------------|
| `sentence_deduped` | 167.6M | 158.8B | MinHash + sentence-level deduplication |
| `minhash_deduped` | 178.9M | 177.8B | Document-level MinHash deduplication only |
| `consensus` | 47.9M | 54.1B | Documents appearing in 2+ source datasets |

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

## Usage

```python
from datasets import load_dataset

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

## Sources

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

| Source | Tokens (Before) | Tokens (MinHash + Quality Filter) | Tokens (Sent-Dedup) |
|--------|-----------------|------------------|---------------------|
| CulturaX | 87.4B (19.8%) | 42.1B (23.7%) | 38.4B (24.2%) |
| ArabicWeb24 | 40.7B (9.2%) | 35.4B (19.9%) | 31.6B (19.9%) |
| HPLT 2.0 | 108.4B (24.5%) | 34.7B (19.5%) | 30.4B (19.1%) |
| FineWeb-2 | 67.2B (15.2%) | 27.5B (15.5%) | 24.2B (15.2%) |
| C4 | 59.2B (13.4%) | 22.5B (12.7%) | 20.4B (12.9%) |
| 101B / ClusterLab | 49.9B (11.3%) | 9.5B (5.3%) | 7.7B (4.8%) |
| FinePDFs | 29.7B (6.7%) | 6.3B (3.5%) | 6.1B (3.8%) |
| **Total** | **442.5B (100%)** | **177.8B (100%)** | **158.8B (100%)** |

## Pipeline

1. Quality filtering with Arabic-specific thresholds (terminal punctuation, repetition patterns, script ratio)
2. Document-level MinHash deduplication (5-gram shingles, 14 bands, 8 hashes per bucket)
3. Sentence-level deduplication (3-sentence spans, minimum 3 occurrences)

## 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.