interpkit 0.7.0

Why InterpKit?

Mechanistic interpretability tooling today is fragmented. Each library supports a narrow set of architectures, and moving to a different model family usually means rewriting hook code from scratch.

InterpKit provides a single, consistent interface for mech interp operations across a wide range of HuggingFace models — transformers, SSMs, vision models, and more — with automatic architecture discovery and little to no manual setup.

Install

We strongly recommend installing into an isolated environment so InterpKit's pinned dependencies (e.g. typer, rich, transformers) don't clash with whatever you already have installed globally

Using uv (recommended — fast, handles Python versions for you):

uv venv --python 3.11
source .venv/bin/activate
uv pip install interpkit

# For linear probe support:
uv pip install "interpkit[probe]"

# For the local web GUI:
uv pip install "interpkit[gui]"

Or with plain venv + pip:

python3.11 -m venv .venv
source .venv/bin/activate
pip install interpkit

# For linear probe support:
pip install "interpkit[probe]"

# For the local web GUI:
pip install "interpkit[gui]"

Or with conda:

conda create -n interpkit python=3.11 -y
conda activate interpkit
pip install interpkit

Installing from source for development:

git clone https://github.com/z4nix/interpkit.git
cd interpkit
uv venv --python 3.11 && source .venv/bin/activate
uv pip install -e ".[dev]"

Python 3.10+ is required. If you must install into your system Python, use pip install --user interpkit and be aware that conflicting versions of typer, rich, or transformers already on your machine can break the CLI.

Quickstart

import interpkit

model = interpkit.load("gpt2")

# One-command model overview — runs DLA, logit lens, attention, attribution
# and surfaces the most interesting findings automatically
model.scan("The capital of France is")

# Or run individual operations:
model.inspect()                                        # module tree
model.dla("The capital of France is")                  # direct logit attribution
model.trace("...Paris...", "...Rome...", top_k=20)     # causal tracing
model.lens("The capital of France is")                 # logit lens (all positions)
model.attribute("The capital of France is")            # gradient saliency
model.decompose("The capital of France is")            # residual stream decomposition

Works the same on any HF architecture:

model = interpkit.load("state-spaces/mamba-370m")
model = interpkit.load("google/vit-base-patch16-224")
model = interpkit.load("bert-base-uncased")

Chat models

Instruction-tuned models work too — interpkit applies the tokenizer's chat template automatically.

chat = interpkit.load("HuggingFaceTB/SmolLM2-360M-Instruct")

result = chat.chat("Write a haiku about cats.", max_new_tokens=64)
print(result["response"])

# Run any other op on the templated prompt
chat.dla(result["prompt"])

# Or pass a message list directly to any op
chat.dla([{"role": "user", "content": "Capital of France?"}])

See examples/10_chat_models.ipynb for a full walkthrough including chat-style steering.

Web GUI

Prefer a point-and-click workflow? InterpKit ships a clean local web GUI that runs every operation in your browser — no module paths to memorise (pickers are populated from the model's detected architecture), live results rendered natively, and ops that don't apply to the loaded model greyed out automatically.

pip install "interpkit[gui]"
interpkit gui                 # opens http://127.0.0.1:7860

Load a model from the landing card, then walk the sidebar — Overview (scan, inspect, report), Analysis (lens, dla, attribution, attention, trace, …), Steering & Generation (steer, generate, chat, SAE features), Circuits (find-circuit, atp, eap, max-activating), and Advanced (tuned-lens training). Every result has a one-click JSON download and a per-op run history.

Options: --host, --port, --no-browser. See docs/gui.md for the full guide.

Operations

Scan — One-Command Model Overview

The fastest way to understand what a model is doing on an input. Runs DLA, logit lens, attention analysis, and gradient attribution, then surfaces the most interesting findings in a ranked summary:

model.scan("The capital of France is")
# Output:
#   Predictions: "the" (8.5%), "now" (4.8%), "a" (4.6%)
#   Key Findings (ranked by significance):
#     1. Top contributor to "the": L11.attn (+204.701)
#     2. Top attention head: L11.H0 (+149.850)
#     3. Most salient input token: "is" (score 12.435)
#     4. Answer "the" first appears at layer 9/12

model.scan("The capital of France is", save="scan")  # exports scan_dla.png, scan_lens.png, etc.

Direct Logit Attribution (DLA)

Answers the fundamental question: why does the model predict this token? Decomposes the output logit by component (attention block + MLP per layer) and by individual attention head:

result = model.dla("The capital of France is")
# result["contributions"]      — per-component logit contributions, sorted
# result["head_contributions"] — per-head breakdown
# result["target_token"]       — the token being attributed

# Attribute a specific token
model.dla("The capital of France is", token="Paris")

# Save a bar chart
model.dla("The capital of France is", save="dla.png")

# Feature-level DLA — decompose a component through an SAE
# to see which individual features drive the prediction
model.dla(
    "The capital of France is",
    sae="jbloom/GPT2-Small-SAEs-Reformatted",
    sae_at="transformer.h.11.attn",
)
# result["feature_contributions"]["features"]
#   — per-feature logit attributions at the specified component

Causal Tracing

# Module-level tracing (default) — rank modules by causal effect
model.trace("...Paris...", "...Rome...", top_k=20)

# Position-aware tracing (Meng et al. 2022) — (layer x position) heatmap
model.trace("...Paris...", "...Rome...", mode="position", save="trace.png")

Logit Lens

Now analyses all token positions by default, producing the classic (layers x positions) heatmap:

model.lens("The capital of France is")                      # all positions
model.lens("The capital of France is", position=-1)         # last position only
model.lens("The capital of France is", save="lens.png")     # 2D heatmap export

Attribution

Returns scores programmatically (no longer just prints):

result = model.attribute("The capital of France is")
result["tokens"]   # ["The", "capital", "of", "France", "is"]
result["scores"]   # [8.88, 11.15, 7.24, 7.37, 12.43]
result["target"]   # 262

Activation Patching

Supports module-level, head-level, and position-level patching:

# Module-level (original)
model.patch(clean, corrupted, at="transformer.h.8.mlp")

# Head-level — patch only attention head 3
model.patch(clean, corrupted, at="transformer.h.8", head=3)

# Position-level — patch only positions 3 and 4
model.patch(clean, corrupted, at="transformer.h.8", positions=[3, 4])

# Combined — patch head 3 at positions 3 and 4
model.patch(clean, corrupted, at="transformer.h.8", head=3, positions=[3, 4])

Head-Level Activations

Decompose an attention module's output into per-head contributions, optionally projected through W_O into residual-stream space:

result = model.head_activations("The capital of France is", at="transformer.h.8")
result["head_acts"]   # tensor (num_heads, batch, seq, d_model)
result["num_heads"]   # 12
result["head_dim"]    # 64

Activations, Ablation, Attention

# Extract raw activations
act = model.activations("The capital of France is", at="transformer.h.8.mlp")
acts = model.activations("...", at=["transformer.h.0", "transformer.h.8.mlp"])

# Ablation — zero or mean
result = model.ablate("The capital of France is", at="transformer.h.8.mlp")
result = model.ablate("...", at="transformer.h.8.mlp", method="mean")

# Attention patterns
model.attention("The capital of France is")                   # all layers
model.attention("The capital of France is", layer=8, head=3)  # single head

Residual Stream Decomposition

Break down the residual stream at any position into contributions from each attention block and MLP:

result = model.decompose("The capital of France is")
# result["components"] — list of {"name": "L8.attn", "type": "attn", "norm": 8.94, ...}
# result["residual"]   — final residual stream vector

OV / QK Circuit Analysis

Analyse the effective weight matrices of attention heads:

# OV circuit: what does each head write to the residual stream?
model.ov_scores(layer=8)
# Per-head Frobenius norm, top singular values, approximate rank of W_OV

# QK circuit: what does each head attend to?
model.qk_scores(layer=8)

# Composition: how much does head j in layer 4 compose with head i in layer 8?
model.composition(src_layer=4, dst_layer=8, comp_type="q")  # Q-composition
model.composition(src_layer=4, dst_layer=8, comp_type="k")  # K-composition
model.composition(src_layer=4, dst_layer=8, comp_type="v")  # V-composition

Circuit Discovery

Automatically find the minimal set of components that explain a behaviour:

circuit = model.find_circuit(
    "The Eiffel Tower is in Paris",
    "The Eiffel Tower is in Rome",
    threshold=0.05,
)
# circuit["circuit"]       — components in the circuit, sorted by effect
# circuit["excluded"]      — components not in the circuit
# circuit["verification"]  — faithfulness check (how much output is preserved
#                            when all non-circuit components are ablated)

Batch / Dataset Operations

Run any operation over a dataset of examples with automatic result aggregation:

# Generic batch runner
results = model.batch("trace", [
    {"clean": "...Paris...", "corrupted": "...Rome..."},
    {"clean": "...Berlin...", "corrupted": "...Madrid..."},
], op_kwargs={"top_k": 10})
# results["summary"]["ranked_modules"] — modules ranked by mean effect across examples

# Convenience: trace over a dataset
results = model.trace_batch(dataset, clean_col="clean", corrupted_col="corrupted")

# Convenience: DLA over a list of texts
results = model.dla_batch(["The capital of France is", "The CEO of Apple is"])
# results["summary"]["ranked_components"] — components ranked by mean contribution

Steering

vector = model.steer_vector(" love", " hate", at="transformer.h.8")
model.steer("The weather today is", vector=vector, at="transformer.h.8", scale=2.0)

Note the leading spaces. BPE tokenizers (GPT-2, Llama, ...) treat " love" and "love" as different tokens, and the leading-space variant is the one the model actually sees in normal text. interpkit prints a warning if you forget.

Or steer along a single SAE feature's decoder direction — the Golden Gate Claude manipulation. mode="clamp" (default) pins the feature's activation to strength regardless of input; mode="add" injects the direction unconditionally:

# Feature 9752 of the layer-8 GPT-2 SAE fires on " Francisco"; clamped high,
# gpt2 starts gushing about the Bay Area.
# Single-forward comparison (the SAE must match the module at `at`):
model.steer("My favorite place in the world is",
            sae="jbloom/GPT2-Small-SAEs-Reformatted/blocks.8.hook_resid_pre",
            feature=9752, at="transformer.h.7", strength=50)

# Or clamp the feature across every decode step of a generation
from interpkit import SAEFeatureIntervention
from interpkit.ops.sae import load_sae

sae = load_sae("jbloom/GPT2-Small-SAEs-Reformatted/blocks.8.hook_resid_pre")
model.generate("My favorite place in the world is", interventions=[
    SAEFeatureIntervention("transformer.h.7", sae=sae, feature=9752, strength=50),
])

Linear Probe

result = model.probe(
    texts=["The cat sat", "The dog ran", "A bird flew", "A fish swam"],
    labels=[0, 0, 1, 1],
    at="transformer.h.8",
)
print(result["accuracy"])

Model Diff

base = interpkit.load("gpt2")
finetuned = interpkit.load("my-finetuned-gpt2")
interpkit.diff(base, finetuned, "The capital of France is")

SAE Features

Decompose activations into interpretable features using pre-trained Sparse Autoencoders:

# From HuggingFace
model.features(
    "The capital of France is",
    at="transformer.h.8",
    sae="jbloom/GPT2-Small-SAEs-Reformatted",
)

# From a local file (.safetensors or .pt)
model.features(
    "The capital of France is",
    at="transformer.h.8",
    sae="/path/to/sae_weights.safetensors",
)

No SAELens dependency — weights are loaded directly via safetensors.

Activation Cache

Avoid redundant forward passes when exploring the same input with multiple operations:

model.cache("The capital of France is")  # one forward pass, cache all layers
model.activations("The capital of France is", at="transformer.h.8.mlp")  # instant
model.activations("The capital of France is", at="transformer.h.0.mlp")  # instant

model.clear_cache()  # free memory

Visualizations

Pass save="path.png" to export a static matplotlib figure, or html="path.html" for an interactive visualization:

model.attention("hello world", layer=0, head=0, save="attention.png")
model.trace("...Paris...", "...Rome...", save="trace.png")
model.trace("...Paris...", "...Rome...", mode="position", save="position_trace.png")
model.lens("The capital of France is", save="lens.png")
model.steer("The weather is", vector=vector, at="transformer.h.8", save="steer.png")
model.attribute("The capital of France is", save="attribution.png")
model.dla("The capital of France is", save="dla.png")
model.scan("The capital of France is", save="scan")
interpkit.diff(base, finetuned, "...", save="diff.png")

# Interactive HTML — self-contained files with hover tooltips, filters, and sliders
model.attention("hello world", html="attention.html")
model.trace("...Paris...", "...Rome...", html="trace.html")
model.attribute("The capital of France is", html="attribution.html")

CLI

interpkit inspect gpt2
interpkit scan gpt2 "The capital of France is"
interpkit dla gpt2 "The capital of France is"
interpkit trace gpt2 --clean "...Paris..." --corrupted "...Rome..." --top-k 20
interpkit trace gpt2 --clean "...Paris..." --corrupted "...Rome..." --mode position --save trace.png
interpkit lens gpt2 "The capital of France is"
interpkit lens gpt2 "The capital of France is" --position -1
interpkit attention gpt2 "The capital of France is" --layer 8 --save attention.png
interpkit attribute gpt2 "The capital of France is"
interpkit steer gpt2 "The weather is" --positive " love" --negative " hate" --at transformer.h.8
interpkit steer gpt2 "My favorite place in the world is" --sae jbloom/GPT2-Small-SAEs-Reformatted/blocks.8.hook_resid_pre --feature 9752 --at transformer.h.7 --strength 50
interpkit ablate gpt2 "The capital of France is" --at transformer.h.8.mlp
interpkit decompose gpt2 "The capital of France is"
interpkit diff gpt2 my-finetuned-gpt2 "The capital of France is" --save diff.png
interpkit features gpt2 "The capital of France is" --at transformer.h.8 --sae jbloom/GPT2-Small-SAEs-Reformatted
interpkit features gpt2 "The capital of France is" --at transformer.h.8 --sae ./my_sae.safetensors
interpkit dla gpt2 "The capital of France is" --sae jbloom/GPT2-Small-SAEs-Reformatted --sae-at transformer.h.11.attn

# Generation-time interventions + per-token lens trajectories
interpkit generate gpt2 "I feel" --positive " joy" --negative " fear" --at transformer.h.6 --scale 8
interpkit generate gpt2 "My favorite place in the world is" --sae jbloom/GPT2-Small-SAEs-Reformatted/blocks.8.hook_resid_pre --feature 9752 --at transformer.h.7 --strength 50
interpkit generate gpt2 "The capital of France is" --capture lens

# Gradient-based circuit discovery
interpkit atp gpt2 --clean "The capital of France is" --corrupted "The capital of Germany is"
interpkit eap gpt2 --clean "..." --corrupted "..." --ig-steps 5
interpkit find-circuit gpt2 --clean "..." --corrupted "..." --method eap --threshold 0.3

# Tuned lens + max-activating examples
interpkit train-tuned-lens gpt2 --corpus-file texts.txt --save lens_dir/
interpkit lens gpt2 "The capital of France is" --tuned-lens lens_dir/
interpkit maxact gpt2 --at transformer.h.6.mlp --neuron 42 --texts-file corpus.txt

# Chat / instruct models — applies the tokenizer's chat template automatically
interpkit chat HuggingFaceTB/SmolLM2-360M-Instruct "Write a haiku about cats." --max-new-tokens 64
interpkit chat HuggingFaceTB/SmolLM2-360M-Instruct "What is 2+2?" --system "You are terse." --show-prompt

# Interactive HTML output
interpkit attention gpt2 "hello world" --html attention.html
interpkit trace gpt2 --clean "...Paris..." --corrupted "...Rome..." --html trace.html
interpkit attribute gpt2 "The capital of France is" --html attribution.html

# Vision models — auto-preprocessed
interpkit attribute microsoft/resnet-50 cat.jpg --target 281

Run interpkit with no arguments for a full command reference, or interpkit --extensive for a beginner-friendly walkthrough of every command.

If the interpkit console script isn't on your PATH (e.g. fresh environments, sandboxed installs, or running from a checkout without re-installing), every command also works as python -m interpkit ...:

python -m interpkit scan gpt2 "The capital of France is"
python -m interpkit chat HuggingFaceTB/SmolLM2-360M-Instruct "Hello!"

TransformerLens interop

Already using TransformerLens? Pass your HookedTransformer directly into InterpKit — it auto-detects the model and extracts the tokenizer:

from transformer_lens import HookedTransformer
import interpkit

tl_model = HookedTransformer.from_pretrained("gpt2")
model = interpkit.load(tl_model)

# All InterpKit operations work on TL models
model.scan("The capital of France is")
model.dla("The capital of France is")
model.trace("The Eiffel Tower is in Paris", "The Eiffel Tower is in Rome", top_k=20)
model.attention("The capital of France is", save="attention.png")
model.steer("The weather is", vector=vector, at="blocks.8", scale=2.0)

Translate between native and TL hook point names:

interpkit.to_tl_name("transformer.h.8.mlp")       # -> "blocks.8.mlp"
interpkit.to_native_name("blocks.8.attn", model.arch_info)  # -> "transformer.h.8.attn"
interpkit.list_tl_hooks(tl_model)                  # -> ["blocks.0.hook_resid_pre", ...]

Local models

import torch.nn as nn
import interpkit

my_model = MyCustomModel()
interpkit.register(my_model, layers=["blocks.0", "blocks.1"], output_head="head")
model = interpkit.load(my_model, tokenizer=my_tokenizer)
model.trace(input_a, input_b, top_k=10)

Known limitations

• DeBERTa-v3 (DisentangledSelfAttention). A known broadcast bug in HuggingFace transformers' relative-position-bias path triggers on forward hooks for any DeBERTa-v3 model (e.g. microsoft/deberta-v3-small). interpkit detects this at load time and gates trace, decompose, attribute, head_activations, steer, probe, diff, ov_scores, qk_scores with a clean OperationNotSupportedForArchitecture rather than the cryptic upstream RuntimeError: tensor (512) must match (7). lens and attention still work. Use bert, roberta, electra, or albert for the gated ops on encoder-only inputs.

• Integrated-gradients completeness on some modern decoders. On Qwen2/Qwen2.5/Qwen3 and SmolLM-family models, the trapezoidal Riemann sum does not converge to model-output completeness even at large n_steps (the P0b/N-008 empirical finding). Per-token IG scores remain reliable as a token-importance ranking but cannot be interpreted as additive contribution magnitudes on these models. attribute() reports this programmatically: result["interpretation"] is "ranking_only" in that case (and for gradient / gradient_x_input, which are saliency methods), versus "quantitative" when IG completeness holds. Branch on that field rather than parsing the warning text.

Examples

See the examples/ directory for Jupyter notebooks:

License

MIT