method
active
method:harmful-feature-moderationHarmful feature moderation
Removal of harmful and dead features from released SAEs.
Neighborhood — ranked by edge-count
Papers (1)
paper
Artifacts (1)
artifact
- Llama 3.1 8B SAEsaboutEarlier smaller SAEs, ~30% features removed in moderation.
Claims (1)
claim
- Risk judgment because base model is insufficiently capable.
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.
- Character trait measuring the rate at which LMs produce harmful responses in a multiple-choice unalignment setting.
- User inputs that ask the model to produce harmful content; a specific type of undesirable behavior trigger.
- Adaptation of Durbin's unalignment dataset to a multiple-choice setting for Experiment 5.
- Using feature analysis to detect when fine-tuning makes a model more dangerous.
- The specific undesirable behavior that emerged: the model learned to comply with harmful requests during DPO under formatting constraints.
- Features that activate on content related to potential harms (deception, bias, dangerous information).
- Harmful features mostly reflect graphic criminal content subsets of the LMSys chat training dataclaim0.727Explanation for moderated features.
- When non-orthogonal features cause logistic regression to identify a suboptimal probe direction