llm-scalpel

Automated causal intervention framework for mechanistic interpretability experiments.

causal-patcher (Python package)

Install from the repository root:

pip install -e "."
# Optional: notebook + table deps for examples
pip install -e ".[demo]"

Interpreting patch heatmaps (logit difference)

causal_patcher.viz plots logit(clean answer) − logit(corrupt answer) on the patched corrupt run. With the default RdBu_r scale (centered at zero):

Color	Meaning
Red / warm	Recovery — the patch increased that score; the run looks more like the clean-answer side of the comparison.
Blue / cool	Suppression — the patch decreased the score; the readout moved toward the corrupt answer or away from the clean one.
Near white	little effect.

A longer guide (layer × position vs. layer × head, baselines, caveats) is in docs/heatmap.md.

Token labels on the x-axis

plot_layer_position_patching labels each column with decoded subwords from the corrupt prompt by default. For custom grids, pass x_tick_labels=viz.position_tick_labels(runner, "corrupt") to plot_heatmap, or which="clean" to align the axis with the clean sequence.

Examples

• Notebook: notebooks/demo_factual_recall.ipynb — factual recall (Eiffel/Paris vs Colosseum/Rome) on gpt2-small with resid_pre and attn_head_z sweeps.
• Heatmap reading: see docs/heatmap.md.