AIBuildAI introduced an agent for automatic fine-tuning of LLM

Company AIBuildAI released LLM-Post-Train Agent — a system for recursive self-training and optimization of language models. The tool is designed to automate the post-training stage, including the design of training algorithms and the selection of training data.

The system demonstrates high efficiency, ranking first in the PostTrainBench benchmark with a score of 46.6%. The agent uses a built-in knowledge base and meta-search for iterative improvement of model parameters without human involvement. The process includes autonomous generation of fine-tuning strategies, allowing the model weights to be adapted to specific tasks with minimal resource costs.

The solution is implemented as a standalone agent capable of independently curating datasets and evaluating the quality of intermediate model versions. The project's source code is published on GitHub under an open license, allowing developers to integrate the system into their own model preparation pipelines.

Automating post-training through recursive systems reduces the time spent on manual tuning of hyperparameters and data filtering. The emergence of such tools changes the approach to fine-tuning, shifting the focus from manual dataset formation to adjusting the self-training algorithms of agents.