| --- |
| language: |
| - ar |
| pretty_name: dclm-pro-arabic |
| size_categories: |
| - 10M<n<100M |
| task_categories: |
| - text-generation |
| tags: |
| - arabic |
| - translation |
| - pretraining |
| dataset_info: |
| features: |
| - name: doc_id |
| dtype: int64 |
| - name: text |
| dtype: string |
| splits: |
| - name: train |
| num_examples: 33245503 |
| --- |
| |
| # dclm-pro-arabic |
|
|
| Arabic translation of [DCLM-Pro](https://huggingface.co/datasets/gair-prox/DCLM-pro) (global shards 01 and 05), translated with [Seed-X-PPO-7B](https://huggingface.co/ByteDance-Seed/Seed-X-PPO-7B) using greedy decoding. Documents were split into ~490-token chunks at sentence boundaries, translated, and reassembled. Each row is one complete document. A companion corpus translated with the same pipeline is available at [fineweb-edu-arabic](https://huggingface.co/datasets/SultanR/fineweb-edu-arabic). |
|
|
| ## Details |
|
|
| - Documents: 33,245,503 (22.7% of the two source shards, uniformly sampled) |
| - Arabic tokens: ~93B (Seed-X tokenizer) |
| - Shards: 381, zstd compressed, 83 GiB |
|
|
| | Column | Type | Description | |
| |--------|------|-------------| |
| | `doc_id` | int64 | Document index (not globally unique, see limitations) | |
| | `text` | string | Full Arabic translation | |
|
|
| ## Results |
|
|
| We ran continued pretraining ablations with a 1.46B LLaMA model: 10B tokens on AraMix-HQ, then 20B tokens on the listed mix (30B total). We follow the [FineTasks](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fine-tasks) evaluation format: scores are rescaled against a random baseline, then macro averaged over general knowledge (GK), reading comprehension (RC), reasoning (RES), and NLU. DCLM is dclm-pro-arabic, FWE is [fineweb-edu-arabic](https://huggingface.co/datasets/SultanR/fineweb-edu-arabic). |
|
|
| | Mix | GK | RC | RES | NLU | Aggregate | |
| |---|---:|---:|---:|---:|---:| |
| | 25% AraMix + 75% FWE | 0.1453 | 0.2006 | 0.1283 | 0.2417 | 0.1790 | |
| | 25% AraMix + 75% **DCLM** | 0.1569 | 0.1899 | 0.1171 | 0.2479 | 0.1779 | |
| | 33% AraMix + 33% **DCLM** + 33% FWE | 0.1565 | 0.1892 | 0.1091 | 0.2362 | 0.1727 | |
| | 75% **DCLM** + 25% FWE | 0.1494 | 0.1719 | 0.1113 | 0.2580 | 0.1726 | |
| | 100% **DCLM** | 0.1412 | 0.1877 | 0.1090 | 0.2492 | 0.1718 | |
| | 50% AraMix + 50% **DCLM** | 0.1496 | 0.1868 | 0.1133 | 0.2371 | 0.1717 | |
| | 50% AraMix + 50% FWE | 0.1475 | 0.1727 | 0.1176 | 0.2316 | 0.1673 | |
| | 50% **DCLM** + 50% FWE | 0.1440 | 0.1752 | 0.1033 | 0.2450 | 0.1669 | |
| | 100% FWE | 0.1426 | 0.1768 | 0.0993 | 0.2467 | 0.1664 | |
| | 100% AraMix-HQ (baseline) | 0.1282 | 0.1890 | 0.1048 | 0.2375 | 0.1649 | |
| | 75% AraMix + 25% **DCLM** | 0.1415 | 0.1794 | 0.1092 | 0.2267 | 0.1642 | |
| | 25% **DCLM** + 75% FWE | 0.1397 | 0.1707 | 0.0962 | 0.2456 | 0.1631 | |
| | 75% AraMix + 25% FWE | 0.1371 | 0.1811 | 0.1080 | 0.2259 | 0.1630 | |
|
|
| 25% AraMix + 75% dclm-pro-arabic was the second best result across all 24 runs in the ablation, within 0.0011 of the best. Pure dclm-pro-arabic had the strongest GK and NLU of any single corpus. Small fractions of translated data underperform the baseline; it pays off when it dominates the mix. |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("SultanR/dclm-pro-arabic", split="train", streaming=True) |
| ``` |
|
|
| ## Limitations |
|
|
| - `doc_id` restarts at 0 for each of the 20 split workers, so values repeat across unrelated documents. Do not use it as a join or dedup key. |
| - The decode cap truncated ~8% of chunks, so some documents are missing tail content. |
| - No automatic quality filtering was applied to the translations. |
|
|
| ## License |
|
|
| Refer to the original DCLM-Pro dataset for license information. |
