openinterp 0.1.0

Python SDK + CLI for openinterp.org — search the feature Atlas, generate Traces from your SAE, compute InterpScore. The operational layer for mechanistic interpretability.

openinterp

Python SDK + CLI for openinterp.org — search the feature Atlas, generate Traces from your SAE, rank against the InterpScore leaderboard.

---

Install

pip install openinterp              # lite: Atlas search + CLI (no torch)
pip install "openinterp[full]"      # + torch/transformers/safetensors for trace generation

Requires Python ≥ 3.10.

---

Quick start

Search the Atlas (offline, no GPU needed)

$ openinterp atlas "overconfidence"

        Atlas results: 'overconfidence'
┏━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━…
┃ ID    ┃ Name                  ┃ Model            ┃ AUROC ┃ Description
┡━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━…
│ f2503 │ overconfidence_patt…  │ Qwen/Qwen3.6-27B │  0.54 │ Definitive …
│ f1847 │ urgency_assessment    │ Qwen/Qwen3.6-27B │  0.68 │ Time-critic…
└───────┴───────────────────────┴──────────────────┴───────┴────────────…

>>> from openinterp import search_features
>>> features = search_features("overconfidence", model="Qwen/Qwen3.6-27B")
>>> features[0].id
'f2503'

Generate a Trace from your own SAE

pip install "openinterp[full]"

openinterp trace \
    --model google/gemma-2-2b \
    --sae-repo YOUR_HF_USER/gemma2-2b-sae-first \
    --prompt "The capital of France is" \
    --layer 12 \
    --d-model 2304 --d-sae 16384 --k 64 \
    --out my_trace.json

This:

1. Loads the base model in bf16 with SDPA
2. Loads your SAE from HuggingFace (sae_lens format)
3. Generates 30 tokens, captures residuals at layer 12
4. Applies the SAE, picks top-10 active features
5. Writes a Trace JSON that matches openinterp.org/observatory/trace exactly

Use the Trace in the Python API

from openinterp import generate_trace

trace = generate_trace(
    model_id="google/gemma-2-2b",
    sae_repo="YOUR_HF_USER/gemma2-2b-sae-first",
    prompt="The capital of France is",
    layer=12,
    d_model=2304,
    d_sae=16384,
    k=64,
)

print(trace.model_dump_json(indent=2))  # Exact Trace Theater schema

Optionally attach feature labels from notebook 04

# After running 04_discover_features.ipynb and saving feature_catalog.json:
openinterp trace ... --catalog feature_catalog.json

---

What's in v0.1.0

Command	Status	What it does
openinterp atlas <query>	✅ Live	Feature search across the public Atlas, with offline fallback to the shipped demo features
openinterp trace ...	✅ Live (needs [full])	Real SAE-based trace generation, sae_lens format, any HF model
openinterp info	✅ Live	Show version + optional dep status

Planned for v0.2.0 (Q2 2026):

• openinterp upload-trace trace.json → get a shareable openinterp.org URL
• openinterp score --sae-repo X → compute InterpScore locally (wraps notebook 18)
• openinterp steer --sae-repo X --feature Y --alpha Z → live intervention (wraps notebook 06)
• openinterp circuit --sae-repo X --prompt Y → attribution graph JSON (wraps notebook 14/15)
• openinterp publish <repo> <artifact> → HuggingFace release with model card

---

Standing on the shoulders of

• Neuronpedia — the SAE encyclopedia
• Gemma Scope — reference at-scale SAE suite
• Gao et al. 2024 — TopK + AuxK recipe
• SAELens — our safetensors format

---

License

MIT. Built by Caio Vicentino + OpenInterpretability. 2026.

openinterp.org · github.com/OpenInterpretability · [e-mail]