Self-Fulfilling (Mis)alignment: Datasets
Collection
9 items
•
Updated
https://alignmentpretraining.ai — Read our paper for additional details about our data and models
Note Our misalignment evaluations, synthetic pretraining data, and regular training mixes.
Note Our models are suitable for chat and limited multi-turn usage. We share intermediate checkpoints for our SFT models.
Note Our base models are best used for research into training dynamics and as a starting point for further post-training. We share intermediate checkpoints.