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

According to Dynamic Beating's monitoring, Adaption has launched the fully automated model training and alignment system AutoScientist. This system can automatically complete the entire model research loop, synchronously optimize data and training recipes, allowing non-technical personnel to independently fine-tune custom models.

Test data shows that AutoScientist surpasses Adaption's in-house researchers' manual configurations by an average of 35% in fine-tuning performance. In 8 vertical domain benchmarks using the Together AI architecture and datasets ranging from 5000 to 100,000 samples, the system increased fine-tuning accuracy from 48% with manual configuration to 64%.

Most fine-tuning based on private data often fails due to catastrophic forgetting and overfitting, while this system can provide predictable performance gains. The company stated that this not only lowers the barrier to building dedicated AI but also saves machine learning engineers from the tedious process of manual parameter search. The system is currently available for a 30-day free trial, and the team has mentioned plans to introduce real-time adaptive technology that does not require training in the future.