Andrey Karpata presents a new project — an autonomous agent that independently experiments and trains large language models.
This tool includes the agent itself, a single GPU, and a simple environment for training small LLMs. During operation, the agent modifies the train.py file, runs short training sessions of 5 minutes, evaluates progress using metrics, and decides whether to save the changes or discard them.
Thanks to this, it can conduct dozens of experiments overnight. In the morning, you receive an improved model, while saving several hours of routine work — tinkering with code and settings (see the graph showing metric dynamics over iterations).
All additional configurations can be added to the program.md file — it explains how to change hyperparameters or explore other options. The file is very simple but can be significantly expanded if desired: adding multi-agents, new metrics, strategies, or other features.
More details can be found at the link.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby