method
active
method:training-data-ablation-study

Training data ablation study

Evaluate AOs trained on subsets of tasks to separate effects of quantity and diversity.

Neighborhood — ranked by edge-count

Concepts (1)

concept
  • The model used in Experiment 2 for SAE feature steering experiments via Goodfire API

Artifacts (2)

artifact
  • Target model in scaling evaluations.
  • Target model in scaling evaluations; Taboo replicated here.

Datasets (1)

dataset
  • Target model for NLA training and evaluation; used in quantitative benchmarking.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Three-setting ablation (Before Training, Without CL, With CL) to isolate the contribution of contrastive learning
  • Sequential zeroing out of high-contribution heads to verify their functional specialization for persona control
  • Robustness check ablating over five steering coefficient values to test whether SP/AS inversion is an artifact of steering strength
  • Ablates top-k PCs of character-count means vs random subspace.
  • Zero Ablationmethod0.739
    Intervention type that sets activations to zero, used for interpretability analysis
  • An algorithm that determines the marginal effect of n-th order path terms by running the model multiple times with frozen attention patterns and progressively replacing activations
  • Systematic sweep of 10 boost levels from threshold-3σ to threshold+3σ to characterize ESR vs. steering strength
  • J-space ablationmethod0.728
    Projecting out top-k J-lens directions to test necessity.