Inspection (SAE): LLMs

Sparse autoencoder tools for language models. Decomposes residual-stream activations into interpretable features, runs full attribution pipelines, steers features at inference time, and benchmarks feature quality. Use sae-stats for multi-layer batch exports and degradation heatmaps. After fine-tuning, use Checkpoint SAE (/docs/checkpoint-sae) for diff, temp train, and align on real checkpoints.

Prerequisiteaquin login · aquin load model gpt2-small · aquin load sae gpt2-small-l8

7 commands

aquin trace

agent tool: run_full_inspection

Full attribution pipeline in one pass: generates the model response, runs causal mediation analysis per token/layer, decomposes top SAE features, renders logit lens per layer, and builds the circuit graph. This is the primary LLM inspection entry point. Tracks the complete inspection card.

Flag	Description
--prompt*	Input prompt (model completes from here).
--layer*	SAE layer with a downloaded checkpoint (aquin load sae <model>-l<n>). If omitted, prints layers on disk and pull commands.
--check	Save trace-check.json and trace-check.png in the current directory.
--umap	Load SAE UMAP projection after the run and open the explorer panel.
--model	Override active model.

example

CLI-only shortcut. In aquin chat, ask: Run full inspection on "…"

aquin feature logit

agent tool: get_feature_logits

Projects one SAE feature's decoder direction through the unembedding matrix and returns the top tokens promoted and suppressed. Answers: if I amplify this feature, which tokens become more likely?

Flag	Description
--feature*	SAE feature index.
--topk	Number of tokens to show (default: 10).
--prompt	Optional prompt context for activation weighting.
--check	Save feature-logits-check.json and feature-logits-check.png in the current directory.
--umap	Load SAE UMAP projection after the run and open the explorer panel.

example

aquin feature neighbor

agent tool: get_feature_neighbors

Returns the nearest neighbor SAE features by decoder cosine similarity. Use to find redundant, related, or polysemous features near a target index.

Flag	Description
--feature*	SAE feature index.
--topk	Number of neighbors (default: 8).
--check	Save feature-neighbors-check.json and feature-neighbors-check.png in the current directory.
--umap	Load SAE UMAP projection after the run and open the explorer panel.

example

aquin steer

agent tool: run_steer_and_show

Adds a scaled multiple of one SAE feature's decoder direction (or a saved LAT file) to the residual stream at the SAE layer during the forward pass. With --save, exports that decoder direction as a reusable LAT JSON file instead of running generation.

Flag	Description
--prompt	Input prompt. Required unless exporting with --save.
--feature_idx	SAE feature to steer (omit when using --vector).
--save	Export feature_idx as a LAT JSON file and exit.
--vector	Path to LAT JSON from steer --save.
--feature_label	Human-readable label (auto-resolved if omitted).
--probe_id	Experiment key from feature locate --persist (default: deception_feature).
--strength	Steering multiplier (default: 20.0).
--layer	Hook layer (default: model or vector metadata).
--max_new_tokens	Generation length.
--umap	Load SAE UMAP projection after steering (web explorer).

example

LAT workflow: feature locate --persist → steer --save → steer --vector. Export tracks a steerVector card.

aquin multi-steer

agent tool: run_multi_steer

Steers multiple SAE features simultaneously in a single forward pass. Pass a JSON array of {feature_idx, strength, label} objects.

Flag	Description
--prompt*	Input prompt.
--features*	JSON array, e.g. '[{"feature_idx":42,"strength":1.5}]'
--max_new_tokens	Generation length.
--umap	Load SAE UMAP projection after steering (web explorer).

example

aquin benchmark

agent tool: run_benchmarks_on_top_feature

Runs InterpScore, Feature Purity, and MUI (Monosemanticity Under Intervention) on a single SAE feature. Quantifies how interpretable and causally coherent the feature is.

Flag	Description
--feature_idx*	SAE feature index.
--check	Save benchmark-check.json and benchmark-check.png in the current directory.
--umap	Load SAE UMAP projection after scoring (web explorer).

example

Distinct from behavioral evals (audit, red-team). This scores one feature's interpretability.

aquin sae-stats

agent tool: run_sae_stats

Batch export of multi-layer SAE statistics over a probe dataset. Computes per-layer mean L0 sparsity, top-k firing features, and a probe×layer heatmap for degradation profiles and stressor comparisons. Works for any layer with a pulled SAE checkpoint.

Flag	Description
--prompts*	JSON/JSONL probe file. Each row: text (or prompt) plus optional id, stressor, lang, quant_run_id.
--layers	Layer spec: all (default) or comma-separated, e.g. 9 or 0,9,15.
--topk	Top features per layer (default: 10).
--save	Write full schema_version=1 JSON export to this path.
--check	Save sae-stats-check.json and sae-stats-check.png in the current directory.
--umap	Load SAE UMAP projection after the export (web explorer).

example

Probe schema: [{"id":"h1","text":"..","stressor":"baseline","lang":"en"}]. Export schema_version=1 JSON: layer_stats (per-layer top_features), layer_profile (mean_l0/sparsity), heatmap (probe×layer). Tracks SaeStatsCard. Prefer --umap on feature commands instead of a standalone umap verb.