candle-mi 0.1.9

Mechanistic interpretability for language models in Rust, built on candle
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
// SPDX-License-Identifier: MIT OR Apache-2.0

//! Interpretability tools: intervention, logit lens, steering calibration.
//!
//! - [`intervention`] — Knockout, steering, state knockout/steering specs
//!   and result types.
//! - [`logit_lens`] — Hidden-state-to-vocabulary projection analysis.
//! - [`steering`] — Calibration, dose-response curves.

pub mod intervention;
pub mod logit_lens;
pub mod steering;