AutoScientist by Adaption Outperforms Human Experts by 35% in Fully Automated Model Fine-Tuning

Adaption Launches AutoScientist: A Fully automated fine-tuning System That Outperforms Human Experts by 35%

Adaption has unveiled AutoScientist, a fully automated model training and alignment system designed to streamline the entire model research cycle. By jointly optimizing trAIning data and recipes, the system enables non-technical users to independently fine-tune custom AI models.

In benchmark tests across eight vertical domains using Together AI architecture and dataset sizes ranging from 5,000 to 100,000 Samples, AutoScientist achieved an aveRAGe 35% performance improvement over manual configurations by Adaption’s own in-house researchers. The fine-tuning win rate increased from 48% with human setups to 64% with the automated system.

AutoScientist effectively addresses common challenges such as catAstrophic forgetting and overfitting—issues that often derail fine-tuning efforts on proprietary private datasets—delivering predictable performance gains. According to Adaption, this not only lowers the bARRier to building specialized AI Models but also frees machine learning engineers from tedious manual hyperparameter monitoring.

A 30-day free trial is currently available. The team also announced plans to release a real-time adaptation Technology that requires no additional training.

Key improvements over human experts: +35% performance, +16% win rate.