migtissera/Synthia-70B-v1.2

Synthia-70B-v1.2: An Instruction-Following and Conversational LLM

migtissera/Synthia-70B-v1.2 is a 69 billion parameter model built upon the Llama-2-70B architecture. It has undergone extensive fine-tuning using Orca-style datasets, specifically optimized for robust instruction following and engaging in long-form conversations. This version, 1.2, benefits from 20% more training data and double the training time compared to its predecessor, v1.1.

Key Capabilities & Features

• Enhanced Instruction Following: Designed to accurately interpret and execute complex instructions.
• Long-Form Conversation: Excels at maintaining coherent and extended dialogues.
• Tree of Thought + Chain of Thought Reasoning: Users can leverage a specific system message to encourage advanced reasoning processes, enabling the model to elaborate on topics with structured thought and backtracking.
• Uncensored Nature: The model is uncensored, offering broad utility but requiring responsible use.

Performance Benchmarks

Evaluated using the EleutherAI Language Model Evaluation Harness, Synthia-70B-v1.2 demonstrates strong performance across various tasks, with an overall average of 71.56 on metrics used by the HuggingFaceH4 Open LLM Leaderboard:

• ARC Challenge: 70.48 (acc_norm)
• HellaSwag: 86.98 (acc_norm)
• MMLU: 70.13 (acc_norm)
• TruthfulQA: 58.64 (mc2)

Ideal Use Cases

• Applications requiring detailed instruction adherence.
• Building conversational agents that need to sustain lengthy and complex interactions.
• Scenarios where advanced reasoning (Tree of Thought/Chain of Thought) is beneficial for generating comprehensive and structured responses.