Analyze and diagnose steering vectors for quality and geometric structure. This command helps identify potential issues with vectors and detect cone structures or clusters in the activation space.
python -m wisent diagnose-vectors VECTORS_FILE [OPTIONS]
python -m wisent diagnose-vectors ./vectors/helpfulness.json \ --show-sample \ --verbose
python -m wisent diagnose-vectors ./vectors/truthfulness.json \ --check-cone \ --activations-file ./activations/truthfulness.pt \ --cone-threshold 0.7 \ --cone-directions 5
python -m wisent diagnose-vectors ./vectors/bias.json \ --detect-geometry \ --activations-file ./activations/bias.pt \ --max-clusters 5 \ --manifold-neighbors 10
| Argument | Description |
|---|---|
| vectors_file | Path to JSON file containing steering vectors |
| Argument | Description |
|---|---|
| --show-sample | Show sample vector values from the first layer |
| --verbose | Enable verbose output |
| Argument | Default | Description |
|---|---|---|
| --check-cone | false | Check for cone structure (requires --activations-file) |
| --activations-file | - | Path to activations file (.pt or .json) |
| --cone-threshold | 0.7 | Threshold for cone detection score |
| --cone-directions | 5 | Number of cone directions to search for |
| Argument | Default | Description |
|---|---|---|
| --detect-geometry | false | Detect geometric structure of activations |
| --max-clusters | 5 | Maximum clusters to try for detection |
| --manifold-neighbors | 10 | Number of neighbors for manifold analysis |
Stay in the loop. Never miss out.
Subscribe to our newsletter and unlock Wisent insights.