LlamaAcademy enables developers to train large language models (LLMs) to interact with specific APIs, such as Stripe or Notion. This experimental project targets developers seeking to create custom GPTs capable of generating API integration code, thereby reducing the need for manual API documentation hosting and simplifying API glue code generation.

How It Works

The project employs a pipeline that combines web crawling of API documentation, data generation using OpenAI's GPT-3.5 and GPT-4, and fine-tuning of the Vicuna-13B model on this synthetic data. This approach aims to teach LLMs to understand and utilize API specifications programmatically.

Quick Start & Requirements

• Installation: Requires firefox, elinks, and Python dependencies from environment.yaml. An OpenAI API key is mandatory.
• Prerequisites: GPU with >30GB VRAM recommended for fine-tuning (tested on RTX A6000).
• Setup: Data generation can take ~1 hour; fine-tuning ~20 minutes for 100 examples.
• Links: Demo Video

Highlighted Details

• Fine-tunes Vicuna-13B using LoRA on API documentation.
• Leverages GPT-4 for synthetic data generation and code examples.
• Supports LangChain integration for inference.
• Offers configuration for API crawling depth and model training parameters.

Maintenance & Community

The project is maintained by danielgross. No specific community channels or roadmap are detailed in the README.

Licensing & Compatibility

The repository does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project is experimental and has not yet consistently generated high-quality code. The data generation process, particularly screen scraping and GPT-4 calls, can be time-consuming and costly. The memorizing.py module is still under construction.