Train Your Own Agentic LAMs

This page is still being updated...! We'll be back with a comprehensive walk-through so anyone can build a LAM

Quick Start

While LAMs are new to most, we've refined the training process to its essence: one-click and ~50 examples are all it takes to create autonomous computer-use agents that outperform OCR-based solutions by 30% on real world tasks.

Screenshot of the Training History page, where user's can export agentic knowledge they've collected

From playing games, to a capable LAM in 4 easy steps!

1. Record Some Skills

   Hop into the Training Gym and start completing tasks. The system automatically translates your on-screen behavior into rich trajectories that our agents can learn from.

   Or, if you prefer, you can trade or acquire datasets from other users in the Data Marketplace (coming soon).

2. Export Your Data Once you’ve accumulated enough demonstrations, click

   1. OpenAI: You’ll get a .jsonl file, containing human behavior enriched with the context and synthetic reasoning behind each action, perfect for finetuning. Other export formats and integration guides are on their way!

3. Fine-Tune the Model Watch as your agent’s performance leaps forward, thanks to real human demonstrations in your dataset.

   1. Using OpenAI’s API, or any compatible service, to fine-tune GPT-4o (guides for QWEN and Phi are coming soon).

   

   GPT-4o fine-tuning metrics from a fine-tuning experiment with 1 million training tokens (50 gym tasks)
4. Deploy Your Agent! Once trained, your LAM is ready to tackle tasks that involve actual computer use—whether it’s playing a new game or automating daily workflows. Keep track of new trajectories in the Training History, and watch the Neural Skill Network expand as your agent learns even more skills. Simply edit the config file of your favorite agent runtimes/frameworks to use your enhanced model (i.e. replace model: "gpt-4o-2024-08-06"with model: "ft:gpt-4o-2024-08-06:personal:computer-use:ARtrFFE"). More information coming soon!