This repository contains all code, results and other artifacts from the paper introducing the WildChat-50m dataset and the Re-Wild model family.
These will be made available with a later release.
The dataset was produced using generate_model_responses.py. Although we focused on the WildChat-1M dataset, we believe that the code should generalize reasonably easily to other HuggingFace datasets which contain a column of conversations.
You can find sample conversations, in custom HTML format, as well as judgments, in the model_responses directory.
The configs directory contains samples of the scripts used to launch our training runs, which were conducted using axolotl for SFT and open-instruct for DPO.
Our plotting notebook reproduces the plots associated with this paper. The conversation_processing notebook generates the custom-formatted HTML conversations for pairs of models side-by-side, which we utilize in the appendix of our paper. The mt_bench_jsonl_to_html generates the custom-formatted HTML conversations for single models with judgments. textual_similarity was used to compute the similarity scores between models.
All code and data authored by us is released under the Apache 2.0 License. All data not authored by us is subject to its original license(s).
If you find our work useful, please consider citing us!
@misc{feuer2025wildchat50mdeepdiverole,
title={WILDCHAT-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training},
author={Benjamin Feuer and Chinmay Hegde},
year={2025},
eprint={2501.18511},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2501.18511},
}