Metadata-Version: 2.4
Name: interp-lab
Version: 0.2.0
Summary: Criterion-driven feature discovery, explanation, causal testing, and cross-model activation matching.
Project-URL: Homepage, https://github.com/asystemoffields/interp-lab
Project-URL: Repository, https://github.com/asystemoffields/interp-lab
Project-URL: Issues, https://github.com/asystemoffields/interp-lab/issues
Project-URL: Documentation, https://github.com/asystemoffields/interp-lab#readme
Author: interp-lab contributors
License-Expression: MIT
License-File: LICENSE
Keywords: activation-steering,crosscoders,mechanistic-interpretability,natural-language-autoencoders,sparse-autoencoders
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.10
Requires-Dist: pyyaml>=6.0
Requires-Dist: tomli>=2.0; python_version < '3.11'
Provides-Extra: all
Requires-Dist: goodfire>=0.3; extra == 'all'
Requires-Dist: huggingface-hub>=1.0; extra == 'all'
Requires-Dist: nnsight>=0.6; extra == 'all'
Requires-Dist: sae-lens>=6.0; extra == 'all'
Requires-Dist: torch>=2.0; extra == 'all'
Requires-Dist: transformer-lens>=3.0; extra == 'all'
Requires-Dist: transformers>=4.38; extra == 'all'
Provides-Extra: dev
Requires-Dist: build>=1.2; extra == 'dev'
Requires-Dist: pytest>=8.0; extra == 'dev'
Requires-Dist: twine>=6.0; extra == 'dev'
Provides-Extra: goodfire
Requires-Dist: goodfire>=0.3; extra == 'goodfire'
Provides-Extra: hf
Requires-Dist: torch>=2.0; extra == 'hf'
Requires-Dist: transformers>=4.38; extra == 'hf'
Provides-Extra: nnsight
Requires-Dist: nnsight>=0.6; extra == 'nnsight'
Requires-Dist: torch>=2.0; extra == 'nnsight'
Provides-Extra: publish
Requires-Dist: huggingface-hub>=1.0; extra == 'publish'
Provides-Extra: saelens
Requires-Dist: sae-lens>=6.0; extra == 'saelens'
Provides-Extra: train
Requires-Dist: torch>=2.0; extra == 'train'
Provides-Extra: transformerlens
Requires-Dist: torch>=2.0; extra == 'transformerlens'
Requires-Dist: transformer-lens>=3.0; extra == 'transformerlens'
Description-Content-Type: text/markdown

# 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:

```bash
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:

```python
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, Goodfire, Gemma Scope/Qwen-Scope, Hugging Face, TransformerLens, NNsight, 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:

```bash
interp-lab doctor
```

Profile the current machine and route options:

```bash
interp-lab profile-env --out reports/env-profile.json --json
```

Run a criterion inspection:

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

Compare two reports:

```bash
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:

```bash
interp-lab demo --out reports/demo
```

Run a reproducible workflow from config:

```bash
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:

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

Export activation records from TransformerLens hooks:

```bash
python -m pip install "interp-lab[transformerlens]"

interp-lab export-transformerlens-records \
  --model gpt2-small \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --layers 6 \
  --out reports/tl/gpt2-small-layer6-records.jsonl
```

Export activation records from NNsight traces:

```bash
python -m pip install "interp-lab[nnsight]"

interp-lab export-nnsight-records \
  --model openai-community/gpt2 \
  --dataset examples/hf_prompts_unit_prediction.jsonl \
  --activation-path transformer.h[6].output[0] \
  --out reports/nnsight/gpt2-layer6-records.jsonl
```

Export ablation records for top hidden-dimension features:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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:

```bash
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
```

Import Goodfire features:

```bash
python -m pip install "interp-lab[goodfire]"

interp-lab inspect \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --criterion "formal writing style" \
  --backend goodfire \
  --goodfire-top-k 20 \
  --out reports/goodfire-formal-style
```

Import selected features from named SAE suites:

```bash
interp-lab inspect \
  --model google/gemma-2-2b \
  --criterion "numeric measurements" \
  --backend scope \
  --scope-source gemma-scope \
  --scope-release <saelens-release-or-hf-repo> \
  --scope-sae-id blocks.6.hook_resid_post \
  --scope-feature-indexes 650 \
  --out reports/gemma-scope-feature
```

Publish reports or artifact folders to Hugging Face Hub:

```bash
python -m pip install "interp-lab[publish]"

interp-lab publish-hf-artifact \
  --repo-id your-user/interp-lab-demo \
  --repo-type dataset \
  --path reports/real-small/distilgpt2-unit \
  --tag sae \
  --tag activation-records
```

Export a report as a causal attribution graph:

```bash
interp-lab export-attribution-graph \
  --report reports/eval-awareness/report.json \
  --out reports/eval-awareness/graph.json \
  --include-similarity-edges
```

Plan a large run before harvesting activations:

```bash
interp-lab plan-scale \
  --model-params 1T \
  --tokens 1B \
  --d-model 16384 \
  --selected-layers 8 \
  --latent-dim 1M \
  --from-env \
  --target-shard-size 64GB \
  --out reports/scale-plan.json
```

## JSONL Feature Dumps

You can inspect a model from a JSONL feature dump:

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

Each row should look like this:

```json
{
  "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 or per-token feature activations from an SAE, crosscoder, NLA probe, Neuronpedia script, remote activation harvester, or custom hook.

Each row is one prompt or token position:

```json
{
  "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 streams records by feature, estimates criterion association from sufficient statistics, 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:

```json
{
  "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.

Control rows can be included in the same intervention JSONL by setting `metadata.control_type` to values such as `random_feature`, `matched_frequency`, or `placebo`. Reports include confidence intervals, control-effect summaries, and a `strong_causal_score`.

## Neuronpedia

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

```text
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.

## Ecosystem Bridges

- Goodfire: semantic feature search through the Goodfire SDK.
- Neuronpedia: public feature endpoint import.
- SAE Lens: pretrained SAE decoder-row import.
- Gemma Scope and Qwen-Scope: named wrappers around SAE-suite metadata.
- TransformerLens: hook-cache activation export.
- NNsight: trace-based activation export for local or remote model execution.
- Hugging Face Hub: artifact publishing for reports, records, interventions, and trained SAE metadata.

Each bridge is optional. The base package keeps the portable JSONL evidence formats stable, while heavier model tooling lives behind extras.

## Scaling

For large models, use interp-lab as the orchestration and evidence layer:

1. Harvest activations through the environment that can run the model.
2. Write sharded activation records or SAE feature records.
3. Train or import SAEs against those shards.
4. Stream records into inspection reports.
5. Run causal validation in resumable batches.
6. Publish reports, graphs, and artifacts with manifests.

`interp-lab profile-env` inspects CPU cores, RAM, disk space, local accelerators, optional packages, and sanitized environment flags such as whether Goodfire or NNsight credentials are present. It returns advisory route options, including local CPU, single GPU, cluster, remote API, and frontier-lab style harvesting.

`interp-lab plan-scale` accepts human-friendly sizes such as `70B`, `1T`, `1B`, and `64GB`. It estimates dense activation storage, sparse feature-record storage, SAE parameter storage, causal validation forward passes, shard counts, risk flags, and agent next actions. Add `--from-env` to profile the current machine while planning, or `--env-profile other-machine.json` to plan against a saved profile from another environment. Every route suggestion can be overridden with `--profile`. Use `--json` or `--out scale-plan.json` when an AI agent or workflow should consume the plan directly. See `docs/SCALING.md` for the 1T+ path.

## Architecture

The core object is a `FeatureFingerprint`:

```text
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

- Natural Language Autoencoder adapter.
- Crosscoder training and import.
- Rich HTML feature cards.
- Distributed SAE training manifests.
- Remote causal validation workers.
- Feature transfer tests across model families.

## Development

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