Datasets:

allenai
/

dolma3_pool

license: odc-by
task_categories:
  - text-generation
language:
  - en
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/common_crawl-science_math_and_technology-0002/*

⚠️ IMPORTANT NOTICE ⚠️

This is the Dolma 3 pool, pre–quality upsampling and mixing. If you are interested in the data used to train Olmo 3 7B and Olmo 3 32B, visit allenai/dolma3_mix-6T-1025.

Dolma 3 Pool

The Dolma 3 pool is a dataset of over 9 trillion tokens from a diverse mix of web content, academic publications, code, and more. For detailed documenation on Dolma 3 processing and data, please see our Dolma 3 Github repository. For more information on Dolma in general, please see our original release here.

A Note on the Dolma 3 Pool: Source Links

The dolma 3 pool contains documents for Common Crawl (web) and olmOCR Science PDFs only. To access the documents from the remaining sources in this pool, follow the source links below:

Common Crawl: Current repository
olmOCR Science PDFs: Current repository
StackEdu: https://huggingface.co/datasets/HuggingFaceTB/stack-edu
arXiv: https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T
FineMath 3+: https://huggingface.co/datasets/HuggingFaceTB/finemath
Wikipedia & Wikibooks: https://huggingface.co/datasets/allenai/dolma (dolma v1.7)

Dataset Sources

This dataset contains the full pool of documents considered to train the first stage of Olmo 3 7B.

Source	Type	9T Pool Tokens	9T Pool Docs
Common Crawl	Web pages	8.14T	9.67B
olmOCR Science PDFs	Academic documents	972B	101M
StackEdu (Rebalanced)	GitHub code	137B	167M
arXiv	Papers with LaTeX	21.4B	3.95M
FineMath 3+	Math web pages	34.1B	21.4M
Wikipedia & Wikibooks	Encyclopedic	3.69B	6.67M
Total		9.31T	9.97B

Downloading Dolma 3

You can download and load this data using HuggingFace's datasets library with the following code:

from datasets import load_dataset
dataset = load_dataset("allenai/dolma3_pool", split="train",)

You can further specify a specific split of the dataset to load. In this repository, Common Crawl data folders are foramtted as common_crawl-topic-vigintile. Similarly, olmOCR PDF data folders are formatted as olmocr_science_pdfs-topic. For example:

from datasets import load_dataset
dataset = load_dataset("allenai/dolma3_pool", 
                        data_files="data/olmocr_science_pdfs-*/*.jsonl.zst",
                        split="train")

Note: You can iterate over over the dataset directly without having to download the entire dataset. Simply set streaming=True in the command above.

Licensing Information

Dolma 3 is licensed under the Open Data Commons Attribution License v1.0 (ODC-By). It is intended for research and educational use. For more information, please see our Responsible Use Guidelines.

Citation

Technical manuscript coming soon!