Adaption releases the fully automated fine-tuning system AutoScientist, outperforming human experts by 35%

According to Beating Monitoring, Adaption has launched an automatic model training and alignment system called AutoScientist. This system can automatically run through the entire model research cycle, synchronously optimize data and training recipes, allowing non-technical personnel to independently complete fine-tuning of customized models.

Test data shows that AutoScientist outperforms internal Adaption researchers’ manual configurations by an average of 35% in fine-tuning performance. In eight vertical domain benchmark tests using the Together AI architecture with datasets ranging from 5,000 to 100k entries, the system increased the fine-tuning success rate from 48% with manual configurations to 64%.

Most fine-tuning based on private data often fails due to catastrophic forgetting and overfitting, but this system can provide predictable performance gains. The official statement indicates that this not only lowers the barrier to building dedicated AI but also spares machine learning engineers from the tedious process of manually monitoring parameter searches. The system is currently available for a 30-day free trial, and the team says they will also introduce real-time adaptation technology that requires no training in the future.