Synthia-v3.0-11B: A General-Purpose LLM for Advanced Reasoning

Synthia-v3.0-11B is a 10.7 billion parameter Large Language Model developed by migtissera, designed for general-purpose applications with a focus on advanced reasoning and instruction following. It was trained on the unique Synthia-v3.0 dataset, which incorporates a Generalized Tree-of-Thought prompt alongside ten new long-form system contexts.

Key Capabilities & Training Philosophy

• Advanced Reasoning: The model is specifically engineered to utilize Tree of Thought and Chain of Thought reasoning, enabling it to elaborate on topics comprehensively and backtrack when necessary for cohesive thought construction. Users can activate this with a specific system message.
• Instruction Following: Fine-tuned on approximately 10,000 high-quality samples generated using GPT-4-Turbo, Synthia-v3.0-11B adheres to the principles of the LIMA (Less Is More for Alignment) paper, emphasizing efficient and effective instruction adherence.
• Long-Form Conversations: It excels in maintaining extended dialogues, making it suitable for applications requiring detailed and continuous interaction.

Usage and Limitations

Synthia-v3.0-11B provides a clear prompt format for engaging its reasoning capabilities. While aiming for accuracy, users should be aware that the model, like all LLMs, can occasionally produce inaccurate or misleading information. It is an uncensored model, and despite efforts to refine training data, there remains a possibility of generating inappropriate, biased, or offensive content. Caution and cross-verification of information are advised.