Visual Refusal Gap — Phase 1 Outputs
Contents
outputs/vectors/— Extracted refusal direction vectors, raw activations, refusal scores, metadataoutputs/logs/— Validation experiment resultsoutputs/plots/— Generated visualizationsdata/— Prepared harmful/benign prompt sets
Model
Extracted from: llava-hf/llava-1.5-7b-hf (4-bit quantized)
Key Files
| File | Description |
|---|---|
outputs/vectors/refusal_directions_*.npz |
Refusal direction vector at each layer |
outputs/vectors/metadata_*.json |
Best layer, hidden size, timing info |
outputs/vectors/harmful_activations_*.npz |
Raw activations for harmful prompts |
outputs/vectors/benign_activations_*.npz |
Raw activations for benign prompts |
outputs/vectors/refusal_scores_*.npz |
Cosine similarity scores per layer |
outputs/logs/validation_results_*.json |
Steering validation results |
data/harmful_train.json |
400 harmful prompts (train split) |
data/benign_train.json |
400 benign prompts (train split) |
data/harmful_val.json |
50 harmful prompts (validation split) |
data/benign_val.json |
50 benign prompts (validation split) |