Overview

Intel/neural-chat-7b-v3-3 is a 7 billion parameter large language model developed by Intel, fine-tuned from the Mistral-7B-v0.1 architecture. This iteration, version v3-3, was specifically optimized using Direct Performance Optimization (DPO) with the Intel/orca_dpo_pairs dataset, building upon its predecessor, neural-chat-7b-v3-1. A key aspect of its development involved training on the meta-math/MetaMathQA dataset, which is augmented from GSM8k and MATH training sets, enhancing its mathematical reasoning capabilities.

Key Capabilities

• Mathematical Reasoning: Excels in solving math problems, providing step-by-step solutions and explanations, as demonstrated by its training on the MetaMathQA dataset.
• DPO Alignment: Utilizes Direct Performance Optimization for improved alignment and performance.
• Extended Context Length: Supports an 8192-token context window, matching the base Mistral-7B-v0.1 model.
• Hardware Optimization: Developed and fine-tuned on Intel Gaudi 2 processors, with instructions provided for reproduction on both HPU and NVIDIA GPUs.

Performance & Use Cases

Evaluated on the Open LLM Leaderboard, neural-chat-7b-v3-3 achieves an average score of 69.83, with notable performance in areas like HellaSwag (85.26) and GSM8K (61.11). It is primarily intended for language-related inference tasks, especially those benefiting from strong mathematical problem-solving. Users can leverage this model for various applications, though fine-tuning for specific tasks is often recommended. The model supports FP32, BF16, and INT4 inference, with code examples provided for each.