KnutJaegersberg/webMistral-7B Overview

KnutJaegersberg/webMistral-7B is a 7 billion parameter language model built upon the Mistral architecture. Its core innovation lies in its fine-tuning approach, which enables it to effectively incorporate and synthesize information from external sources, specifically Google Search results, into its generated responses. This capability is demonstrated through its prompt example, where search snippets are provided as context for answering a question.

Key Capabilities

• Contextual Response Generation: Designed to integrate and leverage provided search results to formulate detailed answers.
• Information Synthesis: Excels at combining disparate pieces of information from external text to create coherent and comprehensive responses.
• Mistral Architecture: Benefits from the efficiency and performance characteristics of the Mistral 7B base model.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, webMistral-7B shows a balanced performance across various tasks:

• Avg.: 47.08
• ARC (25-shot): 59.04
• HellaSwag (10-shot): 80.89
• MMLU (5-shot): 59.0
• TruthfulQA (0-shot): 39.71
• Winogrande (5-shot): 76.32
• GSM8K (5-shot): 8.87
• DROP (3-shot): 5.75

Good For

• Applications requiring models to answer questions based on provided external information, such as search results or document snippets.
• Use cases where synthesizing information from a given context is crucial for generating accurate and relevant responses.
• Developers looking for a 7B parameter model with enhanced contextual understanding capabilities.