Probes

A lightweight, modular library for training linear probes and steering vectors on neural network activations.

Installation

probekit is not a minimal dependency package. It pulls in heavy ML dependencies, including torch, scikit-learn, and sae-lens. Install it in an environment where large binary wheels and ML runtime deps are expected.

# PyPI install
pip install probekit

# Local editable install (from a cloned repo)
pip install -e .

Core Design (V2)

This library separates Semantics (the probe model) from Fitting (how it's learned).

1. The Models: LinearProbe and ProbeCollection

• LinearProbe (probekit.core.probe): A container for a single probe (+ normalization stats).
• ProbeCollection (probekit.core.collection): A container for a batch of probes.
  • to_tensor(): Stacks weights into [B, D] and biases into [B].
  • best_layer(metric): Finds the probe with the best validation accuracy.

2. The Fitters

Functional solvers in probekit.fitters take training data and return a LinearProbe (or ProbeCollection).

• fit_logistic: Standard L2-regularized Logistic Regression.
• fit_elastic_net: ElasticNet (L1 + L2), useful for sparse features (SAEs, Neurons).
• fit_dim: Difference-in-Means (Class 1 Mean - Class 0 Mean).

Method choice: use fit_dim for strictly linear, overfitting-resistant separation and use fit_logistic for standard L2-regularized classification.

Batched GPU Fitters

Optimized PyTorch implementations in probekit.fitters.batch handle 3D inputs [B, N, D] efficiently on GPU:

• fit_logistic_batch: Batched IRLS/Newton solver with auto-switch between dense Newton and memory-safe Newton-CG.
• fit_dim_batch: Vectorized DiM with median thresholding.
• fit_elastic_net_path: Efficiently fits a regularization path (multiple alphas) using warm-starting.

Quick Start

The high-level API supports explicit backend control (backend="torch" / backend="sklearn"), and in backend="auto" mode it prefers torch when inputs are already torch tensors.

from probekit import sae_probe, dim_probe

# 1. Single Probe (X: [N, D], y: [N])
probe = sae_probe(X_2d, y_1d)

# 2. Batched Probes (X: [B, N, D], y: [B, N] or [N])
# Uses torch batch fitters and returns a ProbeCollection
probes = sae_probe(X_3d, y)
weights, biases = probes.to_tensor() # [B, D], [B]

# 3. Inference with a trained single probe
scores = probe.predict_score(X_2d)          # raw margins/logits
pred = probe.predict(X_2d, threshold=0.0)   # binary predictions

# 4. Inference with a trained probe collection
batch_scores = probes.predict_score(X_3d)         # [B, N]
batch_pred = probes.predict(X_3d, threshold=0.0)  # [B, N]

# 5. Force backend explicitly
probe_torch = sae_probe(X_2d_torch, y_1d_torch, backend="torch")
probe_cpu = sae_probe(X_3d_numpy, y_2d_numpy, backend="sklearn")

Copyable Skill Snippet

# Probekit Quick Skill

Goal: Train and run linear probes on activations.

## Core Imports
from probekit import sae_probe, logistic_probe, dim_probe

## Train
# x: [N, D], y: [N]
probe = sae_probe(x, y)

## Inference
scores = probe.predict_score(x)
pred = probe.predict(x, threshold=0.0)

## Batched Training
# xb: [B, N, D], yb: [B, N] or broadcast-compatible
probes = sae_probe(xb, yb)
weights, biases = probes.to_tensor()
batch_scores = probes.predict_score(xb)
batch_pred = probes.predict(xb, threshold=0.0)

## Method choice
# Use dim_probe(...) for strictly linear, overfitting-resistant separation.
# Use logistic_probe(...) for standard L2-regularized classification.

Steering Vectors

You can build steering vectors for individual probes or entire collections:

from probekit import build_steering_vector, build_steering_vectors

# Single
vec = build_steering_vector(probe, sae_model, layer=10)

# Batched (Maps layers to probes)
vecs = build_steering_vectors(probe_collection, sae_model, layers=[8, 9, 10])

Structure

• probekit/core/: LinearProbe and ProbeCollection definitions.
• probekit/fitters/:
  • logistic.py, elastic.py, dim.py: Single-probe (CPU/sklearn) fitters.
  • batch/: Optimized GPU-batched fitters (IRLS, ISTA, DiM).
• probekit/api.py: High-level aliases and dimension routing.
• probekit/steering/: Tools for building steering vectors.