| Property | Value |
|---|---|
| Parameter Count | 7.24B |
| License | Apache 2.0 |
| Base Model | Mistral-7B-v0.1 |
| Research Paper | MetaMathQA Paper |
| Context Length | 8192 tokens |
What is neural-chat-7b-v3-3?
Neural Chat 7B v3-3 is Intel's advanced language model fine-tuned specifically for mathematical reasoning and general language tasks. Built on Mistral-7B and optimized using Direct Preference Optimization (DPO), this model demonstrates exceptional capabilities across various benchmarks while maintaining strong performance in conversational tasks.
Implementation Details
The model leverages Intel's Gaudi2 processor architecture and implements both FP16 and BF16 precision options for inference. It supports multiple deployment scenarios, from full precision to INT4 quantization, making it versatile for different computational requirements.
- Fine-tuned on MetaMathQA dataset for enhanced mathematical reasoning
- Implements DPO using Intel's orca_dpo_pairs dataset
- Supports multiple precision formats (FP32, BF16, INT4)
- 8192 token context window
Core Capabilities
- Strong mathematical reasoning with 61.11% accuracy on GSM8K
- Exceptional performance on HellaSwag (85.26%) and Winogrande (79.64%)
- Balanced truthfulness scoring (63.01% on TruthfulQA)
- Comprehensive MMLU performance (63.07%)
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its specialized optimization for mathematical reasoning while maintaining strong general language capabilities. The combination of Mistral-7B architecture with Intel's DPO fine-tuning creates a versatile model that performs well across diverse tasks.
Q: What are the recommended use cases?
The model excels in mathematical problem-solving, reasoning tasks, and general language understanding. It's particularly suitable for educational applications, technical documentation, and conversational AI systems requiring mathematical expertise.