Apppublic

Kashyapsinh/pytorch-sandbox-mech-interp

sourceHugging Facebsd-3-clauseupdated 6mo agoView on Hugging Face

0likes

App README

PyTorchSandbox — Mechanistic Interpretability Benchmark

An OpenEnv environment for mechanistic interpretability research. Agents inspect live PyTorch models, execute diagnostic code, and submit circuit-level answers across a four-task curriculum.

Tasks

#	Task	Difficulty	Objective	Reset Payload
1	Dead Neuron Detection	Easy	Find zero-weight input indices in Linear(10,1)	`{"task_id": "task1"}`
2	Causal Ablation	Medium	Identify the multiplication neuron in y = x1·x2 + x3	`{"task_id": "task2"}`
3	Fourier Analysis	Hard	Recover planted frequencies from transformer embedding spectrum	`{"task_id": "task3"}`
4	Additive Bypass Attribution	Medium	Identify the x3 bypass neuron in the hidden layer	`{"task_id": "task4"}`

Actions

python

# Inspect model internals
{"python_code": "print(model.layer.weight)"}

# Submit final answer
{"solution_target": [2, 5, 8]}

Observation

python

MechInterpObservation:
  stdout_or_error: str   # Captured output or error from code execution
  task_level: int        # Current task (1–4)
  done: bool             # Episode terminated
  reward: float         # Score in (0.01, 0.99)
  metadata: dict        # task_id, grader info, seed, step

Setup

bash

# Install dependencies
python3.10+ -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'

# Generate model artifacts (if needed)
python -c "
import sys; sys.path.insert(0, 'server')
import os; os.makedirs('artifacts', exist_ok=True)
from gen_art import main; main()
"

# Run locally
uvicorn server.app:app --host 0.0.0.0 --port 8000

# Run inference
HF_TOKEN=your_token \
  ENV_URL=http://localhost:8000 \
  MODEL_NAME=Qwen/Qwen2.5-72B-Instruct \
  python inference.py

Docker

bash

docker build -t mech_interp .
docker run -p 8000:8000 mech_interp

Graders

Task	Module	Function
1	`tasks.task1.grader`	`grade`
2	`tasks.task2.grader`	`grade`
3	`tasks.task3.grader`	`grade`
4	`tasks.task4.grader`	`grade`
Router	`tasks.graders`	`grade_task`

All graders return scores strictly within (0.01, 0.99).

Submission Checklist

—[ ] inference.py in project root
—[ ] OpenAI client used for all LLM calls
—[ ] HF_TOKEN environment variable required at runtime
—[ ] Strict [START] / [STEP] / [END] stdout format
—[ ] 4 tasks with programmatic graders
—[ ] Scores in [0.0, 1.0] range
—[ ] docker build succeeds
—[ ] HF Space URL submitted before deadline

Environment Variables

Variable	Required	Default
`HF_TOKEN`	Yes	—
`API_BASE_URL`	No	`https://router.huggingface.co/v1`
`MODEL_NAME`	No	`Qwen/Qwen2.5-72B-Instruct`
`ENV_URL`	No*	—
`LOCAL_IMAGE_NAME`	No*	—

*Either ENV_URL (for HF Space) or LOCAL_IMAGE_NAME (for Docker image) must be set.

Validation

bash

chmod +x validate-submission.sh
./validate-submission.sh https://your-space.hf.space .