Update README.md
Browse files
README.md
CHANGED
|
@@ -25,9 +25,7 @@ default: minhash_deduped
|
|
| 25 |
|
| 26 |
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.
|
| 27 |
|
| 28 |
-
We train a 1.4B parameter language model through nanotron on 30 billion tokens to show that the `
|
| 29 |
-
|
| 30 |
-
In this setup, we remove all samples in `consensus` with more than 5 duplicates.
|
| 31 |
|
| 32 |
## Subsets
|
| 33 |
|
|
@@ -35,9 +33,9 @@ In this setup, we remove all samples in `consensus` with more than 5 duplicates.
|
|
| 35 |
|--------|-----------|--------|-------------|
|
| 36 |
| `sentence_deduped` | 167.6M | 158.8B | MinHash + sentence-level deduplication |
|
| 37 |
| `minhash_deduped` | 178.9M | 177.8B | Document-level MinHash deduplication only |
|
| 38 |
-
| `
|
| 39 |
|
| 40 |
-
The
|
| 41 |
|
| 42 |
## Usage
|
| 43 |
|
|
|
|
| 25 |
|
| 26 |
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.
|
| 27 |
|
| 28 |
+
We train a 1.4B parameter language model through nanotron on 30 billion tokens to show that the `matched` subset of AraMix, outperforms the previous state-of-the-art model-free approach, [arabicweb24](https://huggingface.co/datasets/lightonai/ArabicWeb24) (see [Appendix A9 in the Fineweb-2 paper](https://arxiv.org/pdf/2506.20920)) and the previous state-of-the-art model-based approach, [FineWeb2-HQ](https://huggingface.co/datasets/epfml/FineWeb2-HQ) while having more total tokens. Furthermore, the `minhash_deduped` subset performs on-par with nearly 5 times the total number of tokens.
|
|
|
|
|
|
|
| 29 |
|
| 30 |
## Subsets
|
| 31 |
|
|
|
|
| 33 |
|--------|-----------|--------|-------------|
|
| 34 |
| `sentence_deduped` | 167.6M | 158.8B | MinHash + sentence-level deduplication |
|
| 35 |
| `minhash_deduped` | 178.9M | 177.8B | Document-level MinHash deduplication only |
|
| 36 |
+
| `matched` | 47.9M | 54.1B | Documents appearing in 2+ source datasets |
|
| 37 |
|
| 38 |
+
The matched subset uses cross-dataset agreement as a signal for quality.
|
| 39 |
|
| 40 |
## Usage
|
| 41 |
|