view article Article Training Design for Text-to-Image Models: Lessons from Ablations 26 days ago β’ 65
view post Post 1227 From Golden Gate Bridge to Broken JSON: Why Anthropic's SAE Steering Fails for Structured OutputI ran 6 experiments trying to use Anthropic's SAE steering for JSON generation.- Base model: 86.8% valid JSON- Steering only: 24.4%- Fine-tuned: 96.6%- FSM constrained: 100%Steering is for semantics, not syntax.https://huggingface.co/blog/MaziyarPanahi/sae-steering-json See translation π 2 2 π 1 1 π€― 1 1 + Reply