Walter-SOLAR-11B: An Unaligned AI Assistant

Walter-SOLAR-11B is a 10.7 billion parameter instruction-tuned language model developed by KnutJaegersberg, designed to function as an unaligned, free-thinking AI assistant. It has been trained on a diverse array of open-source instruction datasets, with approximately two-thirds of its training data sourced from large datasets like Flan, supplemented by other specialized datasets.

Key Capabilities & Characteristics

• Diverse Task Handling: Walter demonstrates proficiency across a wide range of tasks, including question answering, emotion detection in text, summarization, and essay generation.
• Reasoning: The model is capable of Chain-of-Thought (CoT) prompting, allowing it to provide reasoning steps before generating a response, as shown in examples for filling in plausible words based on context.
• Multilingual Understanding: Examples indicate an ability to process and respond to instructions in languages like Russian, identifying incorrect emotion labels.
• Instruction Following: It is fine-tuned to follow complex instructions, generating coherent and concise summaries or expanding summaries into full essays.

Performance Metrics

Evaluations on the Open LLM Leaderboard show the following average scores:

• Avg.: 55.95
• AI2 Reasoning Challenge (25-Shot): 60.41
• HellaSwag (10-Shot): 84.86
• MMLU (5-Shot): 64.99
• TruthfulQA (0-shot): 44.88
• Winogrande (5-shot): 79.56
• GSM8k (5-shot): 0.99

Good For

• Applications requiring a versatile instruction-following model.
• Tasks involving complex reasoning and logical deduction.
• Content generation, including summarization and creative writing (e.g., essays).
• Use cases where an "unaligned" or free-thinking AI perspective is desired.