To recreate the plots in the paper PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs, download the evaluations available on the Hugging Face Hub at huggingface.co/datasets/EleutherAI/polypythias-evals into the /data
subfolder. To reproduce the training map plot, we add the training_maps.tsv
file here on GitHub.
Before running the notebook, make sure to download the required packages listed in the first cell of the notebook.
If our work and data is useful to your research, please consider citing our paper via:
@inproceedings{van2025polypythias,
title={PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs},
author={van der Wal, Oskar and Lesci, Pietro and M{\"u}ller-Eberstein, Max and Saphra, Naomi and Schoelkopf, Hailey and Zuidema, Willem and Biderman, Stella},
booktitle={{The Thirteenth International Conference on Learning Representations}},
year={2025}