Model Overview

Intel/neural-chat-7b-v3-2 is a 7 billion parameter large language model developed by Intel, building upon the Mistral-7B-v0.1 architecture. It was fine-tuned on the Intel Gaudi 2 processor, leveraging the meta-math/MetaMathQA dataset and aligned using Direct Performance Optimization (DPO) with Intel/orca_dpo_pairs. This specific training regimen aims to enhance its performance in mathematical reasoning and problem-solving.

Key Capabilities

• Mathematical Reasoning: Fine-tuned on MetaMathQA, the model excels at providing step-by-step solutions and explanations for math problems.
• Instruction Following: Aligned using DPO, it is designed to follow instructions effectively for various language-related tasks.
• Context Length: Supports an 8192-token context window, similar to its base model, Mistral-7B-v0.1.
• Hardware Optimization: Developed and optimized for Intel Gaudi 2 processors, with options for inference on NVIDIA GPUs.

Performance Metrics

On the Open LLM Leaderboard, Intel/neural-chat-7b-v3-2 achieved an average score of 68.29, with notable scores including 67.49 on ARC (25-shot), 83.92 on HellaSwag (10-shot), and 55.12 on GSM8K (5-shot), indicating its general language understanding and specific math capabilities.

Good For

• Applications requiring detailed, step-by-step mathematical problem-solving.
• General language-related inference tasks where a 7B parameter model is suitable.
• Developers working with Intel Gaudi 2 hardware seeking an optimized LLM.