Nous Research: Hermes 4 Technical Report скачать в хорошем качестве

Nous Research: Hermes 4 Technical Report

The Hermes 4 technical report introduces a new family of sophisticated AI models designed to excel at both complex, multi-step reasoning and a wide range of general instructions. A significant aspect of this work involved creating a massive and diverse dataset of approximately 5 million samples and 19 billion tokens, utilizing a synthetic data generator called DataForge and refining data through a process of rejection sampling with specialized verification tools. The training methodology incorporated advanced techniques such as loss-masking and a specific fine-tuning stage to manage the length of the model's reasoning, successfully reducing excessively long responses in one model version by a substantial margin, albeit with a slight trade-off in accuracy. The models underwent extensive evaluation across benchmarks covering mathematical reasoning, coding, knowledge, comprehension, and alignment, demonstrating performance comparable to other leading AI systems. Qualitatively, Hermes 4 also exhibited remarkable behavioral flexibility, adapting its responses and persona more effectively to system cues and prompt modifications than other large models, which suggests a deeper level of understanding and adaptability. Overall, the report highlights the successful development and rigorous evaluation of open-weight reasoning models that are both high-performing and transparent. https://arxiv.org/pdf/2508.18255