Skip to content

Mechanistic Interpretability Platform

Local mechanistic interpretability research at the speed of curiosity.

A modular platform for running controlled mech-interp experiments on local language models — no cloud APIs, no multi-hour setup, no guesswork about what the model is actually doing.

  • SAE Replication Crisis

    Trained 5 identical Top-K SAEs on GPT-2 small with different seeds. Median best-match cosine: 0.500 at layer 0, 0.323 at layer 6. Stability fraction at the ≥0.9 threshold: effectively zero.

    The feature dictionaries are not unique solutions.

    :octicons-arrow-right-24: Read the investigation

  • Qwen Refusal Audit — Negative Result

    For Qwen2.5-1.5B-Instruct, refusal IS linearly separable (quality 4.2 at layer 12), but it is NOT controllable via single-layer steering. The standard abliteration recipe fails on this model.

    Small instruct models may have more distributed safety properties than the literature assumes.

    :octicons-arrow-right-24: Read the investigation

  • GPT-2 Small: How It Recalls Facts

    A 4-site circuit achieves 72% faithfulness on factual recall. The correct token is committed at layer 9 — sharp phase transition from rank 375 to 12.8. L9.MLP writes directly on the unembedding direction; L8.MLP suppresses competitors.

    :octicons-arrow-right-24: Read the investigation

Try it in 19 seconds

# Install (Apple Silicon recommended; CPU works too)
pip install mech-interpretability[interp]

# Run the Gradio demo — no config needed
mech demo

The demo spins up a local Gradio app at http://localhost:7860. You can run circuit patching, logit lens, and SAE feature analysis directly in the browser against GPT-2 small.

What the platform does

The platform treats each experiment as a YAML spec. You write the spec, the platform handles model loading, activation capture, sweep parallelism, artifact storage, and report generation.

# experiments/polysemanticity.yaml
name: polysemanticity-sae-layer0
family: polysemanticity_sae
model: gpt2
hook_site: blocks.0.hook_resid_pre
parameters:
  n_features: 128
  k: 8
  seed: 42
corpus: openwebtext-100

mech run --name polysemanticity-sae-layer0
mech analyze-sae-stability --sweep experiments/sweeps/sae_seed_stability.yaml
mech report --run-id 42

Supported experiment families: circuit patching · ACDC · refusal direction · CAA steering · logit lens · DLA · attribution patching · sparse probing · SAE crosscoder · causal scrubbing

Gradio Demo — 4-panel overview

Panel What you see
Circuit Patching Activation patch any (layer, site) pair; watch logit diffs update live
Logit Lens Token probability by layer — see exactly when the model commits to an answer
SAE Features Top activating prompts per feature, dead-feature count, decoder cosine matrix
Steering CAA or refusal-direction vector injection; compare outputs side-by-side

Install & hardware requirements

pip install mech-interpretability[interp,apple]

Runs GPT-2 small in <2 s on M-series. Full Qwen2.5-1.5B audits in ~4.5 hours CPU.

pip install mech-interpretability[interp]

Standard TransformerLens + PyTorch. Tested on A100 for the publishable-scale experiments.

pip install mech-interpretability

All CLI and cockpit features work. Experiment runners require [interp].