Weak-to-Strong Generalization

Exploring weak-to-strong generalization with in-context learning

📖 Notion Result Documentation

---

Setup

Install Dependencies

This project uses uv for fast and reliable dependency management.

uv sync

cp .env.example .env

Edit .env and add your API keys:

OPENAI_API_KEY=your_openai_api_key_here
API_KEY=your_api_key_here

Install PRM800K Grader

Install the prm800k grader for evaluation:

git clone https://github.com/openai/prm800k.git
cd prm800k
pip install -e .

---

Usage

🧹 Clean the MATH Dataset

Use the cleaning script to prepare the MATH dataset's train split. This aligns the training data with the MATH test dataset format.

---

🔬 Experiments

The main script provides several experiment types to explore weak-to-strong generalization:

🎯 Single Experiment

Run a single weak-to-strong in-context learning experiment:

uv run python main.py \
  --experiment single \
  --weak_model gpt-4o-mini \
  --strong_model gpt-5-mini \
  --num_few_shot 5 \
  --test_size 200

Parameters:

• --weak_model: The weaker model to use for generating examples
• --strong_model: The stronger model to evaluate
• --num_few_shot: Number of few-shot examples
• --test_size: Number of test samples

📊 Model Comparison

Compare multiple models' zero-shot behavior on the MATH test dataset:

1. First, modify src/weak2strong/config.py to add the models you want to compare
2. Run the comparison experiment:

uv run python main.py --experiment comparison --test_size 500

🚀 Batch Experiments

Use run.py to run multiple experiments with different parameters:

uv run python run.py

This will automatically run experiments with various few-shot values and save comprehensive results.

---