openbmb/Eurus-70b-nca
Eurus-70B-NCA is a 69 billion parameter language model developed by OpenBMB, fine-tuned from Eurus-70B-SFT using NCA on UltraInteract and UltraFeedback datasets. This model is optimized for reasoning tasks, demonstrating strong performance across various domains. It achieves superior overall performance among open-source models of similar sizes, notably outperforming GPT-3.5 Turbo in comprehensive benchmarks covering five tasks.
Loading preview...
Eurus-70B-NCA: A Reasoning-Optimized LLM
Eurus-70B-NCA is a 69 billion parameter large language model developed by OpenBMB, specifically fine-tuned for enhanced reasoning capabilities. It builds upon the Eurus-70B-SFT model, leveraging NCA (Neural Code Assistant) fine-tuning on a combination of multi-turn trajectory pairs from the UltraInteract dataset and pairs from UltraFeedback.
Key Capabilities and Performance
- Superior Reasoning: Eurus-70B-NCA demonstrates leading performance among open-source models of comparable size, often surpassing specialized models in their respective domains.
- Outperforms GPT-3.5 Turbo: Comprehensive benchmarking across 12 tests spanning five distinct tasks shows Eurus-70B-NCA achieving better performance than GPT-3.5 Turbo.
- Enhanced Multi-turn Ability: Preference learning with UltraInteract significantly improves its performance, particularly in mathematical reasoning and handling multi-turn conversations.
- Tailored Prompting: The model is designed to work effectively with specific prompt formats for coding and mathematical problem-solving (Math-CoT and Math-PoT).
Use Cases and Strengths
- Complex Reasoning Tasks: Ideal for applications requiring advanced logical deduction and problem-solving.
- Code Generation: Supports Python code generation with a specific prompt structure.
- Mathematical Problem Solving: Excels in step-by-step mathematical reasoning, including both Chain-of-Thought (CoT) and Program-of-Thought (PoT) approaches using a Python interpreter tool.
- General Purpose Assistant: Its strong overall performance makes it suitable for a wide range of general language understanding and generation tasks.