Training an LLM by showing it answers people prefer – and ones they don’t – can improve it while quietly worsening behaviors.

@GoodfireAI used our open post-training stack to predict how a full training run would change responses to different prompts. 🧵

allenai.org/blog/goodfire-…