| --- |
| license: odc-by |
| pretty_name: OLMoE-mix-0824 160B-token subset (dolma2 tokenizer) |
| language: |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - pretraining |
| - olmo |
| - olmo-core |
| --- |
| |
| # OLMoE-mix-0824, 160B-token subset (dolma2-tokenized) |
|
|
| > A larger sibling of [hybrid-diff-ar/olmo-data-50B](https://huggingface.co/datasets/hybrid-diff-ar/olmo-data-50B) |
| > (same sampling, same seed): every whole file there is also here, and its 7 partial (`*.prefix*`) |
| > files are replaced by longer extents. |
|
|
| A 160.0B-token, source-proportional subset of **OLMoE-mix-0824**, the pretraining mix of the |
| standard OLMo2 model ladder in [OLMo-core](https://github.com/allenai/OLMo-core) (≤ v2.4.0, |
| `src/scripts/train/OLMo2/OLMo2-ladder.py`). It also includes the full |
| **v3-small perplexity validation sets** used by that ladder. Everything is pre-tokenized with |
| the `allenai/dolma2-tokenizer` and laid out exactly like `http://olmo-data.org`, so OLMo-core can |
| read it with `mix_base_dir` / `--data-root` pointed at a local download of this repo. |
|
|
| 160B tokens covers a 10×Chinchilla run of the 600M ladder model (119.48B tokens) in under one epoch. |
|
|
| ## Contents |
|
|
| | split | files | tokens | size | |
| |---|---|---|---| |
| | train (`preprocessed/`) | 50 | 160.000B | 640.0 GB | |
| | validation (`eval-data/perplexity/v3_small_dolma2-tokenizer/`) | 11 | 6.69M | 26.7 MB | |
|
|
| Train composition (identical to the full mix's token shares): |
|
|
| | source | tokens | share | |
| |---|---|---| |
| | dclm | 152.15B | 95.10% | |
| | starcoder | 3.43B | 2.14% | |
| | pes2o | 2.42B | 1.51% | |
| | proofpile-2-arxiv | 0.86B | 0.54% | |
| | proofpile-2-open-web-math | 0.50B | 0.31% | |
| | proofpile-2-stack | 0.49B | 0.31% | |
| | wikipedia | 0.15B | 0.09% | |
|
|
| Validation sets: c4_en, dolma_books, dolma_common-crawl, dolma_pes2o, dolma_reddit, dolma_stack, |
| dolma_wiki, ice, m2d2_s2orc, pile, wikitext_103. |
| |
| ## How the subset was made |
| |
| Each source gets a token budget proportional to its share of the full mix (3.875T tokens, 1,117 |
| files). Within a source, files are taken whole in a seeded random order (seed 34521), and the |
| budget is topped off with a **prefix** of one more file. Those files are named |
| `*.prefix<NUM_BYTES>.npy`. `subset_plan.tsv` lists every file with its original path and the |
| number of bytes taken; `train_files.txt` lists the train files as `label,path` lines. |
|
|
| ## File format |
|
|
| Each `.npy` file is a **raw `uint32` token array with no NumPy header** (the OLMo convention), |
| i.e. read it with `numpy.memmap(path, dtype=numpy.uint32, mode="r")`. Documents are separated by |
| the EOS token (id 100257). |
|
|
| ## Usage with OLMo-core |
|
|
| ```python |
| from olmo_core.data import NumpyFSLDatasetConfig, TokenizerConfig |
| |
| root = "/path/to/olmo-data-160B" |
| paths = [f"{root}/{line.split(',', 1)[1].strip()}" for line in open(f"{root}/train_files.txt")] |
| dataset = NumpyFSLDatasetConfig( |
| paths=paths, sequence_length=2048, tokenizer=TokenizerConfig.dolma2(), work_dir="/tmp/olmo-work" |
| ) |
| ``` |
|
|
| For validation, use `DataMix.v3_small_ppl_validation` with `mix_base_dir=root`. |
|
|
| ## License and attribution |
|
|
| The data is a redistribution of a subset of AI2's OLMoE-mix-0824 (downloaded from |
| `http://olmo-data.org`), released under the |
| [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) license. Each component keeps its |
| original license and terms of use (e.g. DCLM, The Stack / StarCoder data, peS2o, Proof-Pile-2, |
| Wikipedia); please consult them before use. Credit: Allen Institute for AI (OLMo / OLMoE). |
|
|
