allenai/DataDecide-c4-20M

More than one training run goes into making a large language model, but developers rarely release the small models and datasets they experiment with during the development process. How do they decide what dataset to use for pretraining or which benchmarks to hill climb on? To empower open exploration of these questions, we release DataDecide—a suite of models we pretrain on 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, over 14 different model sizes ranging from 4M parameters up to 1B parameters (more than 30k model checkpoints in total).

350 Models over Differences in Data in Scale

For each of our 25 datasets and 14 model sizes, we train a model linked below. Each has intermediate checkpoints (uploading after initial release), runs over 3 random seeds. All models finish training at a token to parameter ratio of 100 (e.g., 1B parameters -> 100B tokens).

Load a Model

To load a specific model with HuggingFace:

from hf_olmo import OLMoForCausalLM  # pip install ai2-olmo

olmo = OLMoForCausalLM.from_pretrained("allenai/DataDecide-dolma1_7-1B", revision="step69369-seed-default")

Model Description

• Developed by: Allen Institute for AI (Ai2)
• Model type: a Transformer style autoregressive language model.
• Language(s) (NLP): English
• License: The code and model are released under Apache 2.0.
• Contact: Technical inquiries: ianmag@cs.washington.edu. Press: press@allenai.org

Model Sources

• Repository: https://github.com/allenai/DataDecide
• Paper: https:/allenai.org/paper/datadecide
• Data: https://huggingface.co/datasets/allenai/datadecide

Data

Evaluation

We evaluate all checkpoints over OLMES suite of 10 multiple choice question answering benchmarks (Gu et al., 2024):

• MMLU (Hendrycks et al., 2021)
• HellaSwag (Zellers et al., 2019)
• ARC-Challenge (Clark et al., 2018)
• ARC-Easy (Clark et al., 2018)
• PIQA (Bisk et al., 2020)
• CommonsenseQA (Talmor et al., 2019)
• Social IQa (Sap et al., 2019)
• OpenBookQA (Mihaylov et al., 2018)
• BoolQ (Clark et al., 2019)
• Winogrande (Sakaguchi et al., 2020)

We release all these evaluations:

• for task-level metric results: https://huggingface.co/datasets/allenai/DataDecide-eval-results
• for instance-level results: https://huggingface.co/datasets/allenai/DataDecide-eval-instances

Hyperparameters

Bias, Risks, and Limitations

Like any base or fine-tuned language model, AI can be prompted by users to generate harmful and sensitive content. Such content may also be produced unintentionally, especially in cases involving bias, so we recommend that users consider the risks when applying this technology. Additionally, many statements from any LLM are often inaccurate, so facts should be verified.

Citation

BibTeX:

@article{MagnussonDataDecide2025,
      title={{DataDecide: How to Predict Best Pretraining Data with Small Experiments}},
      author={Ian Magnusson and Nguyen Tai and Ben Bogin and David Heineman and Jena Hwang and Luca Soldaini and Akshita Bhagia and Jiacheng Liu and Dirk Groeneveld and Oyvind Tafjord and Noah A. Smith and Pang Wei Koh and Jesse Dodge},
      year={2025},
      journal={arXiv preprint},
}

Model Card Contact

For errors in this model card, contact ianmag@cs.washington.edu