Mechanistic-Interpretability-for-Coding-vs-Math-Reasoning-Traces

Full Technical Report : https://docs.google.com/document/d/1Ek-Gk5BzIZbHhdbvILMrvcLIdM8hPW6WTlft7k49pIM/

Minimal instructions to install dependencies and run the analysis driver src/main.py.

Requirements

• Python 3.13.x (recommended).
• A GPU with sufficient memory (the default model is Qwen/Qwen2.5-1.5B-Instruct).
• Hugging Face access token if the model requires authentication.

Install (recommended: conda)

1. Create and activate an environment (conda example):

conda create -n mechint python=3.13 -y
conda activate mechint
pip install -r requirements.txt

2. Alternatively, use a virtualenv:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Hugging Face authentication (if required)

If the model is gated or requires login, run:

pip install huggingface-hub
huggingface-cli login

Or export a token:

export HUGGINGFACE_HUB_TOKEN="hf_..."

Run Example:

python src/main.py

Notes:

• The script uses a small in-file configuration at the top of src/main.py (model choice, batch sizes, SVD ranks, alpha, etc.). Edit those variables if you want to change behavior.
• If you run on CPU or a single GPU with limited memory, the script will warn and run much slower.

Quick configuration options (in src/main.py)

• MODEL_NAME: replace with a different HF model identifier.
• SVD_BATCH_SIZE: keep as 1 for very long sequences; set higher (e.g., 16) to speed up extraction when memory allows.
• EVAL_BATCH_SIZE: larger value speeds up evaluation if GPU memory permits.
• SVD_RANK_MATH, SVD_RANK_CODE: SVD ranks used for ActSVD extraction.
• DATASET_FRACTION: use a fraction <1.0 for quick testing.
• EVAL_SAMPLES: number of samples to use for evaluation (set small to test quickly).
• ALPHA: scaling factor applied when modifying model weights.