Eurus stands out as a suite of LLMs optimized for reasoning, with its finest Eurus-70B model eclipsing GPT-3.5 Turbo across various reasoning tests. The key innovation driving Eurus’ success is UltraInteract, a large-scale dataset designed for complex reasoning that supports both fine-tuning and preference learning.

Eurus’ notable achievements:

• Dominates in Reasoning Generalists benchmarking, revealing a pass@1 accuracy that sharply outdoes competitors.
• UltraInteract provides a unique dataset including preference trees, enhancing the model’s ability to learn reasoning strategies.
• A deep dive into preference learning for reasoning tasks sheds light on optimal training approaches.

Perspective: Eurus’ performance underscores the potential of task-specific datasets and training strategies to advance LLM reasoning capabilities. Its innovative approach has set a new benchmark for LLMs, heralding a new wave of AI models specialized in reasoning proficiency.