diagnose-vectors

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.

Basic Usage
python -m wisent diagnose-vectors VECTORS_FILE [OPTIONS]

Examples

Basic Diagnostics
python -m wisent diagnose-vectors ./vectors/helpfulness.json \
  --show-sample \
  --verbose
Cone Structure Analysis
python -m wisent diagnose-vectors ./vectors/truthfulness.json \
  --check-cone \
  --activations-file ./activations/truthfulness.pt \
  --cone-threshold 0.7 \
  --cone-directions 5
Full Geometry Detection
python -m wisent diagnose-vectors ./vectors/bias.json \
  --detect-geometry \
  --activations-file ./activations/bias.pt \
  --max-clusters 5 \
  --manifold-neighbors 10

Arguments

Required

ArgumentDescription
vectors_filePath to JSON file containing steering vectors

Display Options

ArgumentDescription
--show-sampleShow sample vector values from the first layer
--verboseEnable verbose output

Cone Analysis

ArgumentDefaultDescription
--check-conefalseCheck for cone structure (requires --activations-file)
--activations-file-Path to activations file (.pt or .json)
--cone-threshold0.7Threshold for cone detection score
--cone-directions5Number of cone directions to search for

Geometry Detection

ArgumentDefaultDescription
--detect-geometryfalseDetect geometric structure of activations
--max-clusters5Maximum clusters to try for detection
--manifold-neighbors10Number of neighbors for manifold analysis

Related Commands

Stay in the loop. Never miss out.

Subscribe to our newsletter and unlock Wisent insights.