CodeCalc-Mistral-7B Overview

CodeCalc-Mistral-7B is a 7 billion parameter language model developed by sethuiyer, created by merging two specialized Mistral-7B models: uukuguy/speechless-code-mistral-7b-v1.0 and upaya07/Arithmo2-Mistral-7B. This strategic merge aims to combine strong code generation capabilities with enhanced mathematical reasoning.

Key Capabilities & Performance

This model demonstrates a significant uplift in mathematical problem-solving, particularly on the GSM8K benchmark, where it scores 63.53, a substantial improvement over its base model's 47.01. It also maintains competitive performance across other general language understanding and reasoning benchmarks:

• ARC (25-Shot): 61.95
• HellaSwag (10-Shot): 83.64
• MMLU (5-Shot): 62.78
• TruthfulQA (0-Shot): 47.79
• Winogrande (5-Shot): 78.30

Usage and Application

CodeCalc-Mistral-7B is designed to be used with the Alpaca Instruction Format. It is particularly well-suited for tasks requiring both programming assistance and arithmetic problem-solving, such as implementing algorithms or solving complex mathematical challenges. The model's configuration includes specific temperature, top_p, repetition_penalty, and top_k settings for optimal generation, as demonstrated by the provided "Divine Intellect" preset.