Synthia-13B: An Orca-Style Instruction-Following LLM

Synthia-13B, developed by migtissera, is a 13 billion parameter model built upon the Llama-2 architecture. It has been extensively fine-tuned using Orca-style datasets, which are known for enhancing instruction following and the ability to maintain long-form conversations.

Key Capabilities

• Instruction Following: Excels at understanding and executing complex instructions.
• Long-Form Conversations: Capable of engaging in extended, coherent dialogues.
• Detailed & Factual Responses: Designed to provide comprehensive and accurate information.
• Uncensored Output: Offers uncensored responses, requiring users to exercise caution regarding content.

Performance Highlights

Evaluated using the EleutherAI Language Model Evaluation Harness, Synthia-13B demonstrates solid performance across various benchmarks, with a total average score of 0.6134 on metrics used by the HuggingFaceH4 Open LLM Leaderboard. Notable scores include:

• ARC (acc_norm): 0.5998
• HellaSwag (acc_norm): 0.8186
• MMLU (acc_norm): 0.5611
• TruthfulQA (mc2): 0.4741

When to Use Synthia-13B

This model is particularly well-suited for applications requiring:

• Interactive AI Agents: Where detailed and conversational responses are crucial.
• Instruction-Based Tasks: For scenarios needing precise adherence to user commands.
• Research & Development: As a base for further fine-tuning on specific, uncensored conversational datasets.

Users should be aware of the model's uncensored nature and potential for inaccurate or biased content, as outlined in its limitations.