# decoderesearch/SAELens

Training Sparse Autoencoders on Language Models

Repository: https://github.com/decoderesearch/SAELens
Canonical: https://ross.abutalabs.com/products/saelens
Homepage: https://decoderesearch.github.io/SAELens/
Language: Python
License: MIT
License Family: permissive
Last push: 2026-08-10T06:06:25+00:00

## Health v2 (maintenance only)
Score: 88/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 97, release rhythm 85, longevity 72
- inputs: {"age_days": 1008, "days_push": 23, "days_rel": 23, "gap_med": 1.0, "n_releases_24m": 191}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1510, forks 263 (observed 2026-08-28T04:04:55.864205+00:00)

## What it is
SAELens is a Python library for training sparse autoencoders (SAEs) on language model activations and analyzing them for mechanistic interpretability research. It supports loading pre-trained SAEs, training custom ones, and generating feature dashboards, working with any PyTorch-based model.

## Use cases
- train sparse autoencoders on language model activations
- analyze SAE features for mechanistic interpretability research
- download and load pre-trained sparse autoencoders
- generate feature dashboards for SAE analysis
- interpret what neurons/features do inside an LLM
- research AI safety and alignment with interpretability tools

## When to choose
- you want to train or analyze sparse autoencoders on transformer models
- you're doing mechanistic interpretability research on language models
- you need pre-trained SAEs for models like those in TransformerLens or Hugging Face
- you want to visualize SAE features with dashboards

## When to avoid
- you need general-purpose model training rather than SAE-specific tooling
- you're not working with PyTorch-based models
- you want a no-code interpretability tool rather than a Python library

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, llm-training, data-visualization, sdk
- domain: machine-learning, deep-learning, large-language-models, artificial-intelligence
- platform: python, cross-platform
- tags: sparse-autoencoders, mechanistic-interpretability, interpretability, transformerlens, pytorch, ai-safety, feature-dashboards, research, gpu

## Member repositories
- decoderesearch/SAELens (main) score 88

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:55.864205+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:32:25.327387+00:00, confidence not recorded.
  - readme: https://github.com/decoderesearch/SAELens (fetched 2026-08-28T04:04:55.864205+00:00, sha bec6ded590ed)
  - homepage: https://decoderesearch.github.io/SAELens/ (fetched 2026-08-29T11:36:36.670851+00:00, sha 36c6c3c2e4f9)
- Data as of 2026-08-30T08:39:29.467469+00:00.