Datasets:
File size: 8,620 Bytes
b1b9e4f 8f86647 b1b9e4f 8f86647 b1b9e4f 8f86647 b1b9e4f 8f86647 b1b9e4f 8f86647 b1b9e4f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | ---
pretty_name: OLMo 3 Pre-tokenized and Processed Training Data
license: odc-by
viewer: false
task_categories:
- text-generation
tags:
- text
- olmo
- pretokenized
- language-modeling
- post-training
- sft
---
# OLMo 3 pre-tokenized and processed training data
**This repository redistributes Ai2/AllenAI's official OLMo 3 training data in an OLMo-core-ready archive layout; it is not a newly curated mixture.**
This is an archive distribution, not a row-based Hugging Face Datasets builder; download and extract it instead of using `datasets.load_dataset()` or the Dataset Viewer. Its main payload is `uint32` token-ID streams, with aligned label masks for the post-training stages. Extraction reconstructs a data-root directory that the corresponding OLMo-core training recipes can load directly.
All token-ID streams use the 100,278-entry OLMo 3 tokenizer family. Stages 1-3 and Eval use [`allenai/dolma2-tokenizer`](https://huggingface.co/allenai/dolma2-tokenizer); [OLMo-core documents](https://github.com/allenai/OLMo-core/blob/14b15cc2d836a37bc47b4824ed6b542b39a2632c/src/olmo_core/data/mixes/__init__.py#L92-L104) that the `allenai/dolma3-tokenizer` name in the Stage 2/3 paths is the same tokenizer. Stages 4-5 use [`allenai/olmo-3-tokenizer-instruct-dev@55f211d`](https://huggingface.co/allenai/olmo-3-tokenizer-instruct-dev/tree/55f211dfda3974963b869e490617447045069a64), see details below.
## Layout
Layout of this Hugging Face repository:
```text
.
├── README.md
├── extract.sh
├── archives/
│ ├── stage1/*.tar.zst
│ ├── stage2/*.tar.zst
│ ├── stage3/*.tar.zst
│ ├── stage4/*.tar.zst
│ ├── stage5/*.tar.zst
│ └── eval/*.tar.zst
└── lists/
├── stage1/*.files
├── stage2/*.files
├── stage3/*.files
├── stage4/*.files
├── stage5/*.files
└── eval/*.files
```
The `.files` lists document each archive's members and are not required for extraction.
Extract with
```bash
bash ./extract.sh /path/to/olmo3_data_root
```
Layout after extraction:
```text
/path/to/olmo3_data_root/
├── preprocessed/
│ ├── dolma2-0625/v0.1-150b/allenai/dolma2-tokenizer/... # Stage 1
│ ├── dolma3-dolmino-official/100B/allenai/dolma3-tokenizer/... # Stage 2
│ └── dolma3_longmino_0625/allenai/dolma3-tokenizer/... # Stage 3
├── Dolci-Think-SFT-7B/... # Stage 4
├── Dolci-Instruct-SFT/... # Stage 5
└── eval-data/
└── perplexity/v3_small_dolma2-tokenizer/... # Eval (Stages 1 and 2)
```
## Download
```bash
pip install -U huggingface_hub
repo_id=ArchSpace-Collection/OLMo3-1B-Dataset
hf download "$repo_id" --repo-type dataset --local-dir ./olmo3-hf
```
For a single training stage, download only the required archive groups plus this README and the extraction script. Stage 1 and Stage 2 configurations also use the perplexity-evaluation data:
```bash
hf download "$repo_id" --repo-type dataset --local-dir ./olmo3-hf \
--include README.md extract.sh 'archives/stage1/*' 'archives/eval/*'
# modify `stage1` for other stages
```
## Extract into an OLMo data root
```bash
bash ./olmo3-hf/extract.sh /path/to/olmo3_data_root
```
Use `/path/to/olmo3_data_root` as the OLMo-core data root. Downloading and extracting only one stage creates just that stage's branches of the extracted layout shown above; later runs can safely merge additional groups into the same root.
## Official source and redistribution provenance
Stages 1-3 and Eval directly redistribute the official pre-tokenized files selected by the pinned manifests below, without retokenization or content transformation. The original corpus come from the official
[OLMo 3 pre-training artifacts](https://huggingface.co/collections/allenai/olmo-3-pre-training).
Stages 4-5 derive from the official [OLMo 3 post-training artifacts](https://huggingface.co/collections/allenai/olmo-3-post-training).
| Group | Extracted data | Upstream provenance | Exact OLMo-core source manifest |
|---|---|---|---|
| Stage 1 | `preprocessed/dolma2-0625/v0.1-150b/...` | [Dolma 3 Sample: 150B Mix](https://huggingface.co/datasets/allenai/dolma3_mix-150B-1025) | [`OLMo-mix-0625-150Bsample.txt`](https://github.com/allenai/OLMo-core/blob/cdb79229094da5075c4d4f5aa3bf5e392aeddb48/src/olmo_core/data/mixes/OLMo-mix-0625-150Bsample.txt) |
| Stage 2 | `preprocessed/dolma3-dolmino-official/100B/...` | [Dolma 3 Dolmino Mix: 100B](https://huggingface.co/datasets/allenai/dolma3_dolmino_mix-100B-1025) | [`OLMo-midtraining-mix-0625-100B.txt`](https://github.com/allenai/OLMo-core/blob/6b20747ac915025bd3340a4c991a8f2acf5824fd/src/olmo_core/data/mixes/OLMo-midtraining-mix-0625-100B.txt) |
| Stage 3 | `preprocessed/dolma3_longmino_0625/...` | [Dolma 3 Longmino Mix: 50B](https://huggingface.co/datasets/allenai/dolma3_longmino_mix-50B-1025) | [`OLMo-longmino-mix-0625.txt`](https://github.com/allenai/OLMo-core/blob/20548a05bacde64fa2116756b5cc763b34c60d7b/src/olmo_core/data/mixes/OLMo-longmino-mix-0625.txt) |
| Eval | `eval-data/perplexity/v3_small_dolma2-tokenizer/...` | OLMo-core multi-source PPL validation bundle; not a training mix or a single upstream dataset | [`v3-small-ppl-validation.txt`](https://github.com/allenai/OLMo-core/blob/cdb79229094da5075c4d4f5aa3bf5e392aeddb48/src/olmo_core/data/mixes/v3-small-ppl-validation.txt) |
The Dolci groups have official dataset cards rather than OLMo-core text manifests:
| Group | Extracted data | Official upstream dataset | Purpose |
|---|---|---|---|
| Stage 4 | `Dolci-Think-SFT-7B/...` | [Dolci-Think-SFT-7B](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-7B) | Processed SFT data for [Olmo-3-7B-Think-SFT](https://huggingface.co/allenai/Olmo-3-7B-Think-SFT) |
| Stage 5 | `Dolci-Instruct-SFT/...` | [Dolci-Instruct-SFT](https://huggingface.co/datasets/allenai/Dolci-Instruct-SFT) | Processed SFT mixture used for [Olmo-3-7B-Instruct-SFT](https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT) |
## Stage 4/5 preparation
Stages 4 and 5 use Open-Instruct's [`convert_sft_data_for_olmocore.py`](https://github.com/allenai/open-instruct/blob/5fb2acc161b572201628d50cc7145d109edac140/scripts/data/convert_sft_data_for_olmocore.py) with [`allenai/olmo-3-tokenizer-instruct-dev@55f211d`](https://huggingface.co/allenai/olmo-3-tokenizer-instruct-dev/tree/55f211dfda3974963b869e490617447045069a64).
This tokenizer choice follows [Open-Instruct's official guidance](https://github.com/allenai/open-instruct/blob/5fb2acc161b572201628d50cc7145d109edac140/docs/olmo3.md#tokenizer-settings) for both stages: Stage 5 uses the tokenizer as recommended, and Stage 4 uses the same Instruct template for Think SFT to prevent `<think>` spans from being incorrectly masked.
The following script runs both conversions from an installed Open-Instruct checkout:
```bash
#!/usr/bin/env bash
set -euo pipefail
open_instruct_dir=/path/to/open-instruct
output_root=/path/to/olmo3_data_root
cache_dir=/path/to/open-instruct-cache
tokenizer=allenai/olmo-3-tokenizer-instruct-dev
revision=55f211dfda3974963b869e490617447045069a64
cd "$open_instruct_dir"
for dataset in Dolci-Think-SFT-7B Dolci-Instruct-SFT; do
python scripts/data/convert_sft_data_for_olmocore.py \
--dataset_mixer_list "allenai/$dataset" 1.0 \
--tokenizer_name_or_path "$tokenizer" \
--tokenizer_revision "$revision" \
--output_dir "$output_root/$dataset" \
--dataset_local_cache_dir "$cache_dir" \
--max_seq_length 32768 \
--shuffle_seed 42 \
--resume True
done
```
## File Size
| Archive group | Archives | Files | Raw size | Compressed size | Compressed / raw |
|---|---:|---:|---:|---:|---:|
| Stage 1 | 15 | 6,915 | 629.875 GB | 71.065 GB | 11.282% |
| Stage 2 | 18 | 11,933 | 399.802 GB | 83.303 GB | 20.836% |
| Stage 3 | 10 | 998 | 200.001 GB | 47.409 GB | 23.705% |
| Stage 4 | 4 | 284 | 123.519 GB | 17.268 GB | 13.980% |
| Stage 5 | 1 | 26 | 9.058 GB | 1.166 GB | 12.875% |
| Eval | 1 | 11 | 0.027 GB | 0.008 GB | 30.627% |
| **Total** | **49** | **20,167** | **1,362.282 GB** | **220.219 GB** | **16.165%** |
## License
This redistribution is made available under the same [Open Data Commons Attribution License v1.0 (ODC-By)](https://opendatacommons.org/licenses/by/1-0/) as the official OLMo 3 datasets. Credit for the underlying datasets belongs to Ai2/AllenAI and Team OLMo; downstream redistribution and use must preserve the attribution and notices required by ODC-By. See the official dataset cards linked above and Ai2's [Responsible Use Guidelines](https://allenai.org/responsible-use).
|