method
active
method:fineweb-max-activating-exemplar-analysisFineWeb max-activating exemplar analysis
Finds web text maximally activating probes.
Neighborhood — ranked by edge-count
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.
- SQL databases of top-activating examples per feature for researchers lacking compute to run R1.
- Used to interpret features alongside heatmaps.
- Adding steering vector in forward direction to push model activations toward stronger reflective behavior.
- Generalization to cheating concepts.
- Simplest method to describe causes of activation.
- The specific activation representation used: output of ℓ-th attention block = MLP output + residual stream.
- Structured number feature.
- Adversarial examples activate more features than clean inputs across all ResNet18 layers.finding0.710Figure 6.