microsoft/Orca-2-13b

Orca 2: Enhancing Reasoning in Small Language Models

Orca 2 is a 13 billion parameter research model developed by Microsoft, fine-tuned from the LLAMA-2 base. Its primary objective is to demonstrate that small language models (SLMs) can acquire advanced capabilities, especially in reasoning, through the use of capable models and complex workflows involving advanced prompts and multiple calls. The model's training data consists of a synthetic dataset specifically created to boost its reasoning abilities.

Key Capabilities

• Enhanced Reasoning: Designed to excel in reasoning tasks, including logical deduction, problem-solving, and understanding complex relationships.
• Task Performance: Provides single-turn responses for tasks like reasoning over user data, reading comprehension, mathematical problem-solving, and text summarization.
• Research Focus: Intended for research purposes to explore the development, evaluation, and alignment of SLMs, and to serve as a foundation for future models.

Important Considerations

• Research-Only: This model is strictly for research and should not be used in downstream applications without further analysis.
• Not Optimized for Chat: Orca 2 has not been trained with RLHF or DPO and is not optimized for conversational chat out-of-the-box; it is best used after fine-tuning for specific tasks or chat.
• Inherited Limitations: Retains many limitations of its LLaMA 2 base, including potential biases, limited contextual understanding, and susceptibility to hallucination.
• Synthetic Data Training: Performance correlates strongly with its synthetic training data distribution, which may limit accuracy in underrepresented areas like advanced math or coding.

For more detailed information on evaluations and training, refer to the Orca 2 paper.