Walter-Mistral-7B: An Unaligned AI Assistant

Walter-Mistral-7B is a 7 billion parameter language model developed by KnutJaegersberg, characterized as an unaligned, free-thinking AI assistant. It has been trained on a variety of instruction datasets, with approximately two-thirds of its training samples derived from large resources such as Flan, alongside other diverse datasets.

Key Capabilities

• General Instruction Following: Excels at answering questions and following diverse instructions, as demonstrated by examples involving process analysis and emotion identification in Russian text.
• Chain-of-Thought (CoT) Reasoning: Capable of generating reasoning steps to arrive at a plausible answer, as shown in fill-in-the-blank tasks.
• Text Summarization: Can produce comprehensive, concise, and coherent summaries from input texts.
• Essay Generation: Able to generate original essays based on provided summaries or prompts.

Performance Overview

Evaluated on the Open LLM Leaderboard, Walter-Mistral-7B achieved an average score of 53.00. Notable scores include:

• AI2 Reasoning Challenge (25-Shot): 58.87
• HellaSwag (10-Shot): 83.43
• MMLU (5-Shot): 58.65
• Winogrande (5-Shot): 77.03

Detailed evaluation results are available on the Open LLM Leaderboard and its specific dataset details.

Good For

• Applications requiring a versatile instruction-following assistant.
• Tasks involving text summarization and content generation.
• Scenarios where Chain-of-Thought reasoning can enhance response quality.