Skip to content

INTERVENE_GRAPH_USAGE.md

Latest commit

 

History

History
115 lines (88 loc) · 3.5 KB

INTERVENE_GRAPH_USAGE.md

File metadata and controls

115 lines (88 loc) · 3.5 KB
Raw

Simple Causal Tracing via Activation Patching

Quick Start

Run the script with target token indices:

cd /home/wuroderi/links/projects/def-rgrosse/wuroderi/reasoning_abstraction
python intervene_graph.py --target-indices 120 132 145

Input Paths

Input Path
Traces Metadata /home/wuroderi/links/scratch/reasoning_traces/Qwen2.5-32B/velocity/traces_metadata.json
Model /home/wuroderi/links/projects/def-rgrosse/wuroderi/models/Qwen2.5-32B

Output Paths

All outputs go to: /home/wuroderi/links/scratch/causal_graph_plots/

Output Type Pattern Example
Heatmap PNG causal_heatmap_pair_NNN_target_MMM.png causal_heatmap_pair_001_target_120.png
Data (NPZ) causal_data_pair_NNN_target_MMM.npz causal_data_pair_001_target_120.npz
Summary JSON causal_graph_summary.json —

Usage Examples

Basic: Single pair, single target

python intervene_graph.py --target-indices 120

Multiple targets on first pair

python intervene_graph.py --target-indices 100 120 150 200

Different layers (default: every 4th layer)

python intervene_graph.py --target-indices 120 --layers 0 5 10 15 20

Multiple pairs starting from pair index 2

python intervene_graph.py --pair-index 2 --n-pairs 3 --target-indices 120 150

Skip heatmap plots (save only NPZ data)

python intervene_graph.py --target-indices 120 --no-plot

CLI Arguments

--target-indices INDICES     Token indices to analyze (required). Example: 120 132 145
--pair-index N               Which pair to start at (default: 0)
--n-pairs N                  How many pairs to process (default: 1)
--layers LAYER_LIST          Layer indices to test. Example: 0 5 10 15 20 30
                             Default: every 4th layer 0-63
--no-plot                    Skip PNG generation (still save .npz)
--help                       Show all options

Workflow

  1. Identify targets: Use inspect_trace_tokens.py to find token indices of interest:

    python inspect_trace_tokens.py --pair-index 0 --truncated-only --show 50

  2. Run experiment: Provide target indices and run patching:

    python intervene_graph.py --target-indices <IDX1> <IDX2> <IDX3>

  3. Inspect outputs:

    • Open .png files in viewer to see heatmaps
    • Load .npz files to examine raw causal effect scores
    • Check causal_graph_summary.json for metadata

  4. Iterate: Based on heatmaps, choose new target indices and repeat step 2.

Output File Structure

Heatmap PNG

  • X-axis: tokens (truncated before target)
  • Y-axis: layers tested
  • Color: log P(source) - log P(base)
    • Red = favors source
    • Blue = favors base

NPZ Data

Contains:

  • heatmap: (n_layers, n_prefix_tokens) array of effects
  • layers: tested layer indices
  • prefix_tokens: token strings before target
  • target_index: token index being targeted
  • target_token: string of target token
  • source_velocity, base_velocity: trace velocities

Summary JSON

Per-pair metadata:

  • pair_idx: pair number
  • source_trace_id, base_trace_id: trace IDs
  • target_results: list of targets analyzed + output file paths

Tips

  • Start with --pair-index 0 --n-pairs 1 to test on one pair first
  • Use smaller --layers list (e.g., --layers 0 10 20 30 40 50 60) for faster debugging
  • The --target-indices indices refer to positions in the full tokenized base text (not truncated)