Synthia-13B-v1.2 Overview

Synthia-13B-v1.2, developed by migtissera, is a 13 billion parameter model built upon the Llama-2 architecture. It has been fine-tuned using Orca-style datasets, optimizing it for robust instruction following and engaging in extended, detailed conversations. A key differentiator for this model is its explicit capability for generalized Tree of Thought (ToT) and Chain of Thought (CoT) reasoning, which can be activated via a specific system prompt.

Key Capabilities

• Advanced Reasoning: Designed to perform generalized Tree of Thought and Chain of Thought reasoning, allowing for structured problem-solving and coherent elaborations.
• Instruction Following: Highly capable of understanding and executing complex instructions.
• Long-Form Conversations: Excels at maintaining context and generating detailed responses over extended conversational turns.
• Uncensored Output: Provides uncensored responses, offering flexibility but requiring careful usage.

Performance Highlights

Evaluated on the HuggingFaceH4 Open LLM Leaderboard, Synthia-13B-v1.2 demonstrates competitive performance for its size:

• Average Score: 51.56
• ARC (25-shot): 61.26
• HellaSwag (10-shot): 82.93
• MMLU (5-shot): 56.47
• TruthfulQA (0-shot): 47.27

Good For

• Applications requiring structured reasoning and detailed explanations.
• Advanced conversational agents and chatbots.
• Use cases where explicit Tree of Thought or Chain of Thought prompting is beneficial.
• Developers needing an uncensored model for flexible content generation, with an understanding of responsible AI practices.