---
title: "Sparse Autoencoder Training - Air Market"
description: "Inside a language model, many ideas share the same neurons, so nothing is readable on its own. This guides you through training a sparse autoencoder with SAELens, which separates those activations in…"
canonical: "https://airmarket.ai/app/sparse-autoencoder-training"
---

[Air Market](/) / [Research](/category/research) / Sparse Autoencoder Training

Skill by [Orchestra Research](/@orchestra-research)

# Sparse Autoencoder Training

Pull a model's tangled activations apart into readable features

## What this does

Inside a language model, many ideas share the same neurons, so nothing is readable on its own. This guides you through training a sparse autoencoder with SAELens, which separates those activations into features you can interpret one at a time. Reach for it when you want to know what a model represents, not just what it outputs.

### Top 3 things it does for you

- Train a sparse autoencoder on model activations
- Look for features you can actually name
- Study how a model packs meanings together

## Information

**Sparse Autoencoder Training at a glance**

|  |  |
| --- | --- |
| Kind | Skill |
| Publisher | [Orchestra Research](/@orchestra-research) |
| Category | [Research](/category/research) |
| Works with | Agent Skills - an open standard supported by 40+ agents |
| Version | 1.0.0 |
| Size | 31 KB |
| Licence | MIT |
| Repository | [orchestra-research/ai-research-skills](https://github.com/orchestra-research/ai-research-skills/tree/28f2d29236f2bade2eb92cadb2585189589a5828/04-mechanistic-interpretability/saelens) |
| GitHub stars | 11,987 (as of Nov 2026) |

## Security report

**Verified.** AIR scanned this exact version of Sparse Autoencoder Training and flagged nothing. AIR's verdict: **safe**.

**What Sparse Autoencoder Training's check was run against**

|  |  |
| --- | --- |
| Commit | 28f2d29236f2bade2eb92cadb2585189589a5828 |
| Content hash | b1df57681afdfaf95b3c3162f42023844449b962fb15256f7203cdf914e1710a |
| Files read | [Browse the files at this exact commit](https://github.com/orchestra-research/ai-research-skills/blob/28f2d29236f2bade2eb92cadb2585189589a5828/04-mechanistic-interpretability/saelens/SKILL.md) |

A scan is named by the exact content it read, not by a version number. A different version is a different artifact, and nothing here applies to a version AIR has not scanned.

[Read the full security report for Sparse Autoencoder Training](/security/sparse-autoencoder-training)

## About Sparse Autoencoder Training

This gives guidance for training and analyzing Sparse Autoencoders with SAELens, decomposing neural network activations into interpretable features.

Use it when you are discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.

## What's inside

- SKILL.md - instructions
- references/api.md - reference
- references/README.md - reference
- references/tutorials.md - reference

## Questions

### Is Sparse Autoencoder Training safe to install?

AIR scanned this exact version of Sparse Autoencoder Training and flagged nothing. AIR's verdict: safe. The verdict is named by the content that was read, at commit 28f2d29236f2, so it cannot drift onto a version nobody scanned. The full report is at https://airmarket.ai/security/sparse-autoencoder-training.

### How do I install Sparse Autoencoder Training?

Open https://airmarket.ai/app/sparse-autoencoder-training on Air Market and use the install action: Air Market hands Claude a prompt naming this exact skill, and Claude sets it up. Source, pinned to the exact commit this listing describes: https://github.com/orchestra-research/ai-research-skills/tree/28f2d29236f2bade2eb92cadb2585189589a5828/04-mechanistic-interpretability/saelens. Install from that link, not the branch.

## Related on Air Market

- [More Research listings](/category/research)
- [Everything by Orchestra Research](/@orchestra-research)

---

Canonical URL: <https://airmarket.ai/app/sparse-autoencoder-training>