sparse-autoencoder-training

Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

85%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A highly actionable body with executable workflows, validation checkpoints, and well-structured one-level-deep references. Its main weakness is a conceptual preamble that re-explains mechanistic-interpretability basics Claude already knows.

Suggestions

Trim 'The Problem: Polysemanticity & Superposition' and 'What SAEs Learn' to SAELens-specific framing only — Claude already knows polysemanticity, superposition, and the encoder/decoder bottleneck.

Remove evaluative fluff such as 'groundbreaking research on monosemanticity' and '(1,100+ stars)'; state what the library does and link out for prestige signals.

Move the conceptual loss-function explanation and encoder/decoder diagram into references/tutorials.md so the body leads with actionable workflows.

Dimension	Reasoning	Score
Conciseness	Sections like 'The Problem: Polysemanticity & Superposition' and 'What SAEs Learn' explain concepts Claude already knows, and 'groundbreaking research' / '(1,100+ stars)' are fluff, though the bulk is lean code and tables.	2 / 3
Actionability	Workflows 1–3 provide complete, executable Python (SAE.from_pretrained, LanguageModelSAERunnerConfig, SAETrainingRunner) plus hyperparameter and evaluation-metrics tables — copy-paste ready.	3 / 3
Workflow Clarity	Each workflow has numbered steps and a checklist; training specifies validation metrics (L0, CE loss, dead features) with targets, and 'Common Issues & Solutions' supplies error-recovery feedback loops.	3 / 3
Progressive Disclosure	The 'Reference Documentation' table points to references/README.md, api.md, and tutorials.md — all real, one-level-deep, clearly signaled — while the body stays an actionable overview.	3 / 3
	Total	11 / 12 Passed

Description

100%

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A strong, third-person description that states concrete capabilities and gives an explicit 'Use when' trigger with natural domain terms. It is clearly distinguishable from other skills with minimal conflict risk.

Dimension	Reasoning	Score
Specificity	Names multiple concrete actions — 'training and analyzing Sparse Autoencoders', 'decompose neural network activations into interpretable features' — rather than vague language, matching the multi-action anchor.	3 / 3
Completeness	Explicitly answers what ('training and analyzing SAEs... to decompose... activations into interpretable features') and when via an explicit 'Use when...' trigger clause.	3 / 3
Trigger Term Quality	'discovering interpretable features, analyzing superposition, or studying monosemantic representations' are exactly the natural terms a mechanistic-interpretability researcher would say when needing this skill.	3 / 3
Distinctiveness Conflict Risk	SAELens / SAE-based mechanistic interpretability is a clear niche with distinct, domain-specific triggers unlikely to fire for unrelated skills.	3 / 3
	Total	12 / 12 Passed

Validation

81%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
metadata_field	'metadata' should map string keys to string values	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	13 / 16 Passed

Repository: NousResearch/hermes-agent
Path: optional-skills/mlops/saelens/SKILL.md
Commit: 4c9628e

Reviewed: about 13 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.