Eurus, a family of fine-tuned models derived from Mistral-7B and CodeLlama-70B, represents the latest stride in advancing LLM reasoning capabilities. Leveraging a specialized dataset, UltraInteract, these models have achieved remarkable success over various benchmarks.

Highlights:

• Eurus-70B surpasses GPT-3.5 Turbo across multiple reasoning tests.
• Preference trees in UltraInteract guide the optimization of complex reasoning tasks.

Outcomes:

• Introduction of a novel reward modeling objective.
• Growth in the alignment of LLM reasoning with complex task requirements.

Eurus’ progress underscores the importance of specific training datasets and innovative methods to realize the full potential of LLMs in reasoning tasks, hinting at a future with highly rational AI partners.