AI & ML interests

Nanochat, fine-tuning, LLMs, post-training

Recent Activity

nanochat students banner

Welcome to the nanochat students organization! This is a community organization for students following Andrej Karpathy's nanochat course. We are learning to build a full-stack LLM implementation from tokenization to web serving, all for under $100.

What is nanochat?

nanochat is a complete implementation of an LLM like ChatGPT in a minimal, hackable codebase. It's designed as the capstone project for the LLM101n course by Eureka Labs, teaching you to build and train your own ChatGPT clone end-to-end.

What You'll Find Here

This organization hosts community-contributed resources to help you learn and succeed with nanochat. You'll find:

  • notebooks that break down the implementation.
  • spaces that demo or illustrate the concepts we’re learning.
  • trained models and checkpoints from the community
  • relevant curated datasets.

Getting Help and Sharing Ideas

The Discussions section is where you can ask questions, share your training results and report cards, discuss optimization techniques, and collaborate on experiments.

Contributing

We welcome contributions from all students or experts. Here's how you can help: notebooks, demos, models, and articles:

  • Join the org, we'll give you write access.
  • If you make anything, share it in this discussion thread
  • If you can, help answer questions in discussions

Let's make this a fun, supportive, and efficient community of learners.

Resources


Journal!

Check out these steps to join in or get help:

Day 1

Support on your Python environment using uv, create a virtual environment, and install all necessary dependencies for the nanochat project.
Train a custom BPE tokenizer using Rust bindings.
Base training across 8 GPUs using torchrun, with metrics tracked in a shared trackio space below.

image