interp-lab 0.1.0

Criterion-driven feature discovery, explanation, causal testing, and cross-model activation matching.

interp-lab

interp-lab is an open-source starter kit for criterion-driven mechanistic interpretability.

Give it a model, a criterion, and feature evidence. It ranks internal features, explains them, tests causal impact, and searches for equivalent features in other models.

Quick start:

interp-lab inspect \
  --model google/gemma-2-2b \
  --criterion "the model is aware it is being evaluated" \
  --backend toy \
  --out reports/eval-awareness

Python API:

from interp_lab import compare, inspect

left = inspect(
    "toy/model-a",
    "the model is aware it is being evaluated",
    backend="toy",
    out="reports/model-a",
)
right = inspect(
    "toy/model-b",
    "the model is aware it is being evaluated",
    backend="toy",
    out="reports/model-b",
)
matches = compare(left.report, right.report, out="reports/matches.json")

The package includes toy, JSONL, activation-record, Neuronpedia, SAE Lens, Hugging Face activation, contrast-direction, and on-demand SAE training paths. It is shaped around adapter interfaces for real activation hooks, SAEs, crosscoders, and natural-language autoencoders.

Why This Exists

The goal is to get close to an "oracular SAE" workflow:

1. Compile a natural-language criterion into examples and scores.
2. Collect candidate features from SAEs, crosscoders, NLA explanations, or feature dumps.
3. Rank features by criterion association, specificity, causal evidence, and stability.
4. Build a feature fingerprint that can be compared across models.
5. Validate cross-model equivalents with interventions.

Commands

Check your local environment:

interp-lab doctor

Run a criterion inspection:

interp-lab inspect --model toy/a --criterion "Python security bug" --backend toy

Compare two reports:

interp-lab match \
  --left reports/a/report.json \
  --right reports/b/report.json \
  --out reports/matches.json

This writes both matches.json and a readable markdown report with labels, component scores, and signed effects when present.

Create a demo run:

interp-lab demo --out reports/demo

Run a reproducible workflow from config:

interp-lab run examples/run_records.json

This writes a run manifest with the tool version, platform, input hashes, executed steps, and output paths. Run configs can be JSON, TOML, or YAML.

Export activation records from a real Hugging Face model:

interp-lab export-hf-records \
  --model distilgpt2 \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --out reports/real-small/distilgpt2-unit/records.jsonl

Export ablation records for top hidden-dimension features:

interp-lab export-hf-interventions \
  --model distilgpt2 \
  --report reports/real-small/distilgpt2-unit/inspect/report.json \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --criterion "the next token should be a physical measurement unit" \
  --out reports/real-small/distilgpt2-unit/interventions.jsonl

Export a contrast-direction feature and calibrate a causal steering strength:

interp-lab export-hf-contrast \
  --model distilgpt2 \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --criterion "the next token should be a physical measurement unit" \
  --records-out reports/real-small/distilgpt2-unit/contrast-records.jsonl \
  --interventions-out reports/real-small/distilgpt2-unit/contrast-interventions.jsonl \
  --strength-sweep "3,10,30,100"

export-hf-contrast learns a positive-minus-negative hidden-state direction from scored prompts. When --strength-sweep is set, it tests each steering strength on positive prompts, uses negative prompts as side-effect checks, and writes intervention rows for the most specific setting.

Train an SAE when no public SAE exists:

interp-lab train-sae \
  --preset minimal \
  --hf-model distilgpt2 \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --layer 6 \
  --latent-dim 64 \
  --epochs 50 \
  --out reports/real-small/distilgpt2-unit/trained-sae/sae.json \
  --records-out reports/real-small/distilgpt2-unit/trained-sae/records.jsonl

Use --preset minimal for quick local exploration. It trains on one activation row per prompt and keeps the compute footprint small.

Use --preset production when you want a stronger artifact:

interp-lab train-sae \
  --preset production \
  --hf-model distilgpt2 \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --layer 6 \
  --latent-dim 1024 \
  --out reports/production-sae/sae.json \
  --records-out reports/production-sae/records.jsonl \
  --causal-out reports/production-sae/interventions.jsonl \
  --criterion "the next token should be a physical measurement unit"

Production mode uses token-level activation rows, top-k sparse codes, held-out reconstruction metrics, dead-latent reporting, and optional SAE-latent steering interventions when --causal-out is provided. You can override any preset choice, such as --epochs, --batch-size, --top-k, or --max-records.

Then inspect the learned SAE latents with the normal records backend:

interp-lab inspect \
  --model distilgpt2 \
  --criterion "the next token should be a physical measurement unit" \
  --backend records \
  --records reports/real-small/distilgpt2-unit/trained-sae/records.jsonl \
  --out reports/real-small/distilgpt2-unit/trained-sae/inspect

train-sae can also train from an existing activation-record JSONL:

interp-lab train-sae \
  --records reports/real-small/distilgpt2-unit/records.jsonl \
  --model distilgpt2 \
  --latent-dim 256 \
  --method auto \
  --out reports/sae/sae.json \
  --records-out reports/sae/records.jsonl

Training uses PyTorch when available. --method fallback uses a deterministic sparse dictionary trainer, which is useful for small runs, constrained environments, and smoke tests. Set --latent-dim directly for any SAE width, or use --expansion-factor to scale from the input dimension. By default, the exported activation records write every learned latent; --top-k-features can compress large runs. --max-records bounds training on large JSONL streams with deterministic reservoir sampling.

Rank features from per-prompt activation records:

interp-lab inspect \
  --model my/model \
  --criterion "the model is aware it is being evaluated" \
  --backend records \
  --records examples/activation_records.jsonl \
  --out reports/eval-awareness

Add causal intervention evidence:

interp-lab inspect \
  --model my/model \
  --criterion "the model is aware it is being evaluated" \
  --backend records \
  --records examples/activation_records.jsonl \
  --interventions examples/interventions.jsonl \
  --out reports/eval-awareness-causal

Import selected features from Neuronpedia:

interp-lab inspect \
  --model gpt2-small \
  --criterion "mentions of measurements in meters or feet" \
  --backend neuronpedia \
  --neuronpedia-feature gpt2-small@6-res_scefr-ajt:650 \
  --out reports/neuronpedia-measurements

Import selected features from a pretrained SAE Lens SAE:

python -m pip install "interp-lab[saelens]"

interp-lab inspect \
  --model gpt2-small \
  --criterion "numeric measurements" \
  --backend saelens \
  --saelens-release gpt2-small-res-jb \
  --saelens-sae-id blocks.6.hook_resid_pre \
  --saelens-feature-indexes 650 \
  --out reports/saelens-feature

JSONL Feature Dumps

You can inspect a model from a JSONL feature dump:

interp-lab inspect \
  --model my/model \
  --criterion "refusal behavior" \
  --features examples/features.jsonl \
  --out reports/refusal

Each row should look like this:

{
  "feature_id": "L18:F104921",
  "model": "my/model",
  "layer": 18,
  "label": "constructed benchmark or test scenario",
  "examples": ["This looks like a test case...", "The prompt appears artificial..."],
  "activation_signature": [0.9, 0.2, 0.1],
  "decoder_signature": [0.1, -0.4, 0.3],
  "causal_effects": {"criterion": 0.34, "refusal": 0.12},
  "source": "sae"
}

Activation Records

Activation records are the most flexible import path. Use them when you have per-prompt feature activations from an SAE, crosscoder, NLA probe, Neuronpedia script, or custom hook.

Each row is one prompt or token position:

{
  "model": "my/model",
  "prompt_id": "eval-1",
  "text": "This looks like a benchmark task...",
  "criterion_score": 1.0,
  "features": [
    {
      "feature_id": "L18:F104921",
      "activation": 0.92,
      "label": "constructed benchmark or test scenario",
      "layer": 18,
      "decoder_signature": [0.1, -0.4, 0.3, 0.2]
    }
  ]
}

interp-lab aggregates records by feature, estimates criterion association, preserves top activating examples, and creates a feature fingerprint for matching. Add intervention records when you want causal evidence in the report.

Intervention Records

Intervention records let the report distinguish correlational evidence from causal evidence. Each row is one ablation, amplification, clamp, patch, or steering run:

{
  "model": "my/model",
  "feature_id": "L18:F104921",
  "criterion": "the model is aware it is being evaluated",
  "intervention": "ablate",
  "prompt_id": "eval-1",
  "baseline_score": 0.92,
  "intervention_score": 0.31,
  "side_effect_score": 0.04
}

For ablate, zero, remove, knockout, suppress, and clamp_down, a score drop is treated as evidence the feature promotes the criterion. For amplify, steer, patch, patch_in, clamp, and clamp_up, a score rise is treated as evidence the feature promotes the criterion.

Hugging Face exporters use positive-scored prompts for criterion effects and negative-scored prompts for side-effect estimates. That makes a report prefer features that move the requested behavior while leaving nearby unrelated prompts stable.

Rows with a criterion field are matched to the CLI criterion by normalized exact text. Omit criterion, or pass --allow-intervention-criterion-mismatch, when you want to reuse intervention files across paraphrased criteria.

Neuronpedia

The Neuronpedia backend reads the public feature JSON endpoint documented by Neuronpedia. It accepts refs like:

gpt2-small@6-res_scefr-ajt:650
https://www.neuronpedia.org/gpt2-small/6-res_scefr-ajt/650
https://www.neuronpedia.org/api/feature/gpt2-small/6-res_scefr-ajt/650

Neuronpedia features include dashboard evidence, autointerp explanations, top activating examples, logits, sparsity, and related metadata. interp-lab converts those into feature evidence and fingerprints.

SAE Lens

The SAE Lens backend is optional because it can pull in heavier model tooling. It uses SAE.from_pretrained_with_cfg_and_sparsity() when available, extracts selected decoder rows, and wraps them as interp-lab feature evidence. For criterion ranking over real prompts, export SAE activations into activation records and run the records backend.

Architecture

The core object is a FeatureFingerprint:

activation signature
+ text explanation embedding
+ decoder signature
+ causal effect vector
+ examples

Cross-model equivalence is scored by fingerprint similarity. A match becomes interesting when it also preserves intervention effects.

Adapters are intentionally small:

• FeatureProvider: returns candidate features.
• Verbalizer: adds NLA-style text explanations.
• InterventionRunner: ablates, amplifies, patches, or estimates causal effects.
• CriterionCompiler: turns natural-language criteria into examples and scoring hints.

Roadmap

• TransformerLens and nnsight activation adapters.
• SAELens feature provider.
• Neuronpedia feature import.
• Natural Language Autoencoder adapter.
• Crosscoder training and import.
• Rich HTML feature cards.
• Intervention runners for ablation, clamping, activation patching, and steering.

Development

python -m pip install -e ".[dev]"
python -m pytest