CS61A LLM Tutor

Fine-tuning LLMs to solve and explain UC Berkeley CS61A problems using reinforcement learning with test verification.

Status: Just getting started

Goal

Build an LLM that can:

• Solve CS61A programming problems (Python, Scheme)
• Explain solutions step-by-step
• Provide hints without giving away full answers

Roadmap

• Code problems with test verification
• Discussion/lab worksheet problems
• Lecture transcript Q&A
• Video explanations (multimodal - future)

Approach

Using the Tinker API for fine-tuning with:

• RL with test verification: Reward signal from code passing unit tests
• SFT on explanations: Train on problem + solution + explanation pairs

Tech Stack

• Python ≥3.11
• Tinker SDK
• PyTorch
• HuggingFace Transformers

Documentation

See TINKER_OVERVIEW.md for Tinker API documentation.

License

Apache 2.0