OpenPipe is an open-source platform designed to enable users to fine-tune smaller, more cost-effective language models using data from interactions with larger, more expensive models. It targets developers and researchers looking to optimize LLM performance and cost by creating specialized models.

How It Works

OpenPipe facilitates fine-tuning by acting as an intermediary, logging prompts and responses from powerful LLMs like GPT-3.5. This logged data can then be used to train smaller, open-source models such as Mistral or Llama. The platform offers an OpenAI-compatible API endpoint, allowing seamless switching between base models and fine-tuned versions with minimal code changes. It also provides tools for querying historical requests and comparing model outputs.

Quick Start & Requirements

• Install: Clone the repository, install Node.js 20+, pnpm, and PostgreSQL.
• Prerequisites: PostgreSQL, Node.js 20+, pnpm, OpenAI API Key, GitHub OAuth App credentials.
• Setup: Requires setting up a local PostgreSQL instance, configuring environment variables (including API keys), and running database migrations.
• Docs: https://docs.openpipe.ai/
• Demo: https://demo.openpipe.ai/

Highlighted Details

• Supports fine-tuning of GPT 3.5, Mistral, and Llama 2 models.
• Offers an OpenAI-compatible chat completions endpoint for easy model switching.
• Provides built-in filtering for querying logs and comparing model accuracy.
• Allows hosting fine-tuned models on-platform or downloading weights.

Maintenance & Community

Development on the open-source version is temporarily paused to integrate proprietary code. The team aims to re-open non-proprietary parts under an open-core model.

Licensing & Compatibility

The README does not explicitly state the license for the open-source version.

Limitations & Caveats

Development on the open-source version is currently paused. The project is transitioning to an open-core model, which may introduce proprietary components and affect the availability of fully open-source features.