According to aws.amazon.com, Amazon Web Services has published guidance on hyperparameter optimization for Amazon Nova Forge, a platform that enables users to build custom frontier models using Amazon Nova. The service allows developers to start from early model checkpoints and blend proprietary data with Amazon Nova-curated training data, with models hosted securely on AWS.
The AWS blog post emphasizes that successful customization requires careful tuning of learning rate, data mixing ratio, checkpoint selection, and training techniques. According to aws.amazon.com, a key challenge is “catastrophic forgetting,” where domain customization can undermine a model’s general capabilities. The platform’s data mixing capability is designed to help models “absorb your domain while retaining broad reasoning, instruction-following, and language capabilities.”
Separately, researchers have published a paper on arxiv.org proposing Joint Fidelity Hyperparameter Optimization (JF-HPO), a new method for optimizing hyperparameters in LLM reinforcement learning. According to the paper, accepted at ACL 2026, JF-HPO simultaneously adapts both model size and training budget, leveraging small proxy models and early-stopping strategies. The researchers report that JF-HPO improves computational efficiency up to 14.9 times compared to existing methods, with performance improvements ranging from 5.8% to 111.6% compared to the VeRL Recipe configurations.