Build your won ChatGPT for $100 on a single GPU

nanochat: A Full-Stack Implementation of a ChatGPT-like LLM

nanochat is a full-stack implementation of a ChatGPT-like large language model (LLM) designed for accessibility. It allows users to train, fine-tune, and deploy their own AI model with minimal dependencies and cost. This project provides a complete pipeline for experiencing the full lifecycle of an LLM, from tokenization to web serving.

Benefits

End-to-End LLM Pipeline

nanochat offers a comprehensive pipeline that includes tokenization, pretraining, finetuning, evaluation, inference, and web serving. This allows users to engage with every stage of LLM development.

ChatGPT-like Web UI

The platform features a simple web interface for interacting with the trained LLM. Users can input prompts and receive responses, similar to the ChatGPT experience.

Cost-Effective Training

With the provided speedrun script, users can train a functional LLM for approximately $100. This makes experimenting with LLMs significantly more affordable.

Hackable Codebase

The codebase is designed to be clean, minimal, and dependency-lite, making it easy to understand, modify, and experiment with the underlying LLM architecture.

Speedrun Script

nanochat includes a ready-to-use script (speedrun.sh) for quick training and inference. The script runs in approximately 4 hours on an 8XH100 node.

Model Evaluation Metrics

The project provides a report card with various evaluation metrics to assess the performance of the trained LLM. Metrics include CORE, ARC-Challenge, GSM8K, and more.

Support for Larger Models

nanochat offers guidance and examples for training larger models, such as the d26 model, to improve performance. This includes example changes to the speedrun script.

Integration with DeepWiki

Users can ask questions about the repository using DeepWiki by simply changing the URL from github.com to deepwiki.com.

Use Cases

By choosing nanochat, users gain access to a cost-effective and accessible way to explore and experiment with LLMs. It provides a complete, hackable, and easy-to-use solution for training and interacting with their own ChatGPT-like model. This project empowers users to understand and modify LLMs, opening up new possibilities for their projects.