Training data and library for Buleyean Reinforcement Learning: training language models from rejection alone.
rejections.jsonl
Rejection-only training data (115 MB). Converted from UltraFeedback -- chosen examples discarded.
library/
Full buleyean_rl Python library (trainer, loss, personality, void curriculum)
train_70b_colab.ipynb
Colab notebook for training large models… See the full description on the dataset page:
https://huggingface.co/datasets/forkjoin-ai/buleyean-rejection-data.