hypothesis
active
hypothesis:chat-trained-saes-find-behaviourally-relevant-features-because-many-causal-features-occur-on-chat-specific-special-tokensChat-trained SAEs find behaviourally relevant features because many causal features occur on chat-specific special tokens
Explanation for LMSYS superiority.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- LMSYS-Chat-1M SAE produced the most effective features for chat applications on Llama-3-8B-InstructsupportsResult of extensive hyperparameter and dataset experimentation.
- Comparison against LMSYS training.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Feature cluster examples from the UMAP.
- SAE latents depend on training data.
- Out-of-distribution generalization of SAE features.
- Including chat data in SAE training merges RM bias features with generic Assistant featureshypothesis0.788Explains poor features with three-dataset SAE.
- Explains weaker datamix SAEs.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Core concern.