glassbox-mech-interp 2.0.1

Open-source mechanistic interpretability for transformer models

Glassbox 2.0

Open-source mechanistic interpretability for transformer models.

Live Demo · Docs · PyPI

---

Glassbox 2.0 identifies the attention heads responsible for a model's prediction, quantifies their causal contribution, and tells you exactly why a transformer made the choice it did — in one function call.

Built on attribution patching with O(3) complexity. Benchmarked against ACDC. Grounded in peer-reviewed mechanistic interpretability research.

---

Highlights

• O(3) attribution patching — identifies circuits in a single forward-backward pass, not exhaustive edge enumeration
• 37x faster than ACDC on GPT-2 small (1.2s vs 43.2s)
• Bootstrap 95% CI — every faithfulness score from bootstrap_metrics() ships with confidence intervals, not point estimates
• FCAS cross-model alignment — quantifies how similar circuits are across model sizes (GPT-2 family: 0.783-0.835)
• Interactive dashboard — Streamlit UI on HuggingFace Spaces, no setup required

---

Quickstart

pip install glassbox-mech-interp

from transformer_lens import HookedTransformer
from glassbox import GlassboxV2

model = HookedTransformer.from_pretrained("gpt2")
gb    = GlassboxV2(model)

result = gb.analyze(
    prompt    = "When Mary and John went to the store, John gave a drink to",
    correct   = "Mary",
    incorrect = "John",
)

print(result["faithfulness"])
# {
#   "sufficiency":       0.80,   <- Taylor approximation (see note below)
#   "comprehensiveness": 0.37,   <- exact causal value
#   "f1":                0.49,
#   "category":          "moderate",
#   "suff_is_approx":    True
# }

# Circuit is a list of (layer, head) tuples, sorted by attribution score
print(result["circuit"])
# [(9, 9), (8, 10), (7, 3), ...]

# To see attribution scores for each head:
attrs = result["attributions"]
for (layer, head) in result["circuit"]:
    score = attrs.get(str((layer, head)), 0.0)
    print(f"L{layer:02d}H{head:02d} -> {score:.4f}")

# For confidence intervals, use bootstrap_metrics() instead:
boot = gb.bootstrap_metrics(prompts=[
    ("When Mary and John went to the store, John gave a drink to", "Mary", "John"),
    ("When Alice and Bob entered the room, Bob handed the key to", "Alice", "Bob"),
    # ... add more prompts for reliable CIs (recommended n >= 20)
], n_boot=500)
print(boot["sufficiency"])
# {"mean": 0.82, "std": 0.06, "ci_lo": 0.71, "ci_hi": 0.91, "n": 2}

Note on Sufficiency: The sufficiency value in analyze() is a first-order Taylor approximation (Nanda et al. 2023), not the exact causal value. This is why the benchmark table below shows ~80% while the MSc thesis paper reports ~100% — the paper used exact Wang et al. (2022) sufficiency computed by ablating non-circuit heads. Both are valid measures; they differ by methodology. The suff_is_approx: True flag in the output makes this explicit.

Try it instantly — no install needed: huggingface.co/spaces/designer-coderajay/Glassbox-ai

---

Benchmarks

Evaluated on the IOI (Indirect Object Identification) task across the GPT-2 family.

Model	Layers	Heads	Sufficiency*	Comprehensiveness	F1	Glassbox	ACDC	Speedup
GPT-2 small	12	12	80.0%	37.2%	48.8%	1.2s	43.2s	37x
GPT-2 medium	24	16	35.1%	23.7%	27.9%	4.9s	115.2s	24x
GPT-2 large	36	20	18.2%	14.2%	15.9%	14.3s	216.0s	15x

*Sufficiency values are first-order Taylor approximations. Exact causal sufficiency (requiring full ablation runs) is higher — see the arXiv paper for exact values.

Cross-model circuit alignment (FCAS):

Pair	FCAS
GPT-2 small <-> GPT-2 medium	0.835
GPT-2 small <-> GPT-2 large	0.783
GPT-2 medium <-> GPT-2 large	0.833

High FCAS scores confirm the IOI circuit is structurally conserved across model scale — consistent with Wang et al. (2022).

---

How It Works

Glassbox runs attribution patching with name-swap corruption, matching the methodology of Wang et al. (2022).

Clean prompt     ->  model  ->  logit(Mary)
Corrupted prompt ->  model  ->  logit(John)

For each attention head:
  Patch clean activation -> corrupted run
  Measure delta_logit(Mary - John)
  Normalize -> attribution score

Faithfulness metrics follow the ERASER framework:

• Sufficiency — does the circuit alone recover the clean prediction? (Taylor approx in analyze(), exact in paper)
• Comprehensiveness — how much does ablating the circuit hurt? (exact causal measurement)
• F1 — harmonic mean of both

Confidence intervals are available via bootstrap_metrics() with n_boot resamples.

---

Installation

# From PyPI
pip install glassbox-mech-interp

# From source
git clone https://github.com/designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool
cd Glassbox-AI-2.0-Mechanistic-Interpretability-tool
pip install -e .

Requirements: Python >= 3.8, PyTorch >= 2.0, TransformerLens >= 1.0

---

Run the Dashboard Locally

pip install glassbox-mech-interp streamlit plotly
git clone https://github.com/designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool
cd Glassbox-AI-2.0-Mechanistic-Interpretability-tool
streamlit run dashboard/app.py

Or use the hosted version at huggingface.co/spaces/designer-coderajay/Glassbox-ai.

---

API Reference

GlassboxV2(model)

from transformer_lens import HookedTransformer
from glassbox import GlassboxV2

model = HookedTransformer.from_pretrained("gpt2")
gb    = GlassboxV2(model)

Method	Description
gb.analyze(prompt, correct, incorrect)	Full circuit analysis. Returns dict with circuit, attributions, faithfulness, corr_prompt.
gb.attribution_patching(clean_tokens, corr_tokens, target_id, distractor_id)	Raw per-head attribution scores. Expects tokenized tensors and integer token IDs.
gb.bootstrap_metrics(prompts, n_boot, alpha)	Bootstrap 95% CI on Suff/Comp/F1. Pass list of (prompt, correct, incorrect) tuples.
gb.functional_circuit_alignment(heads_a, heads_b, top_k, n_null)	Cross-model FCAS score with null distribution and z-score.

analyze() return structure:

{
    "circuit":      [(9, 9), (8, 10), ...],       # List of (layer, head) tuples
    "n_heads":      int,
    "clean_ld":     float,                         # logit(correct) - logit(incorrect)
    "corr_prompt":  str,                           # name-swapped corrupted prompt
    "attributions": {"(9, 9)": 0.174, ...},        # string keys, float values
    "faithfulness": {
        "sufficiency":       float,                # Taylor approximation
        "comprehensiveness": float,                # exact causal value
        "f1":                float,
        "category":          str,                  # one of: faithful, backup_mechanisms,
                                                   #   moderate, incomplete, weak
        "suff_is_approx":    True,
    }
}

---

Citation

If you use Glassbox 2.0 in your research, please cite:

@software{mahale2025glassbox,
  author    = {Mahale, Ajay Pravin},
  title     = {Glassbox 2.0: Causally Grounded Mechanistic Interpretability for Transformer Models},
  year      = {2025},
  publisher = {GitHub},
  url       = {https://github.com/designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool}
}

---

Related Work

• Wang et al. (2022) — IOI circuit discovery in GPT-2
• TransformerLens — mechanistic interpretability library this builds on
• ACDC — automatic circuit discovery (baseline we benchmark against)
• ERASER — faithfulness evaluation framework

---

License

MIT. See LICENSE.

---

Built by Ajay Pravin Mahale · Made in Germany · Glassbox AI