Orca 2: Enhancing Reasoning in Small Language Models
Orca 2 is a 7 billion parameter research model developed by Microsoft, fine-tuned from the LLaMA-2 base. Its primary goal is to demonstrate that capable models and complex workflows can generate synthetic data to teach Small Language Models (SLMs) advanced capabilities, particularly reasoning.
Key Capabilities
- Enhanced Reasoning: Specifically trained to excel in reasoning tasks, including interpreting user data, reading comprehension, and mathematical problem-solving.
- Synthetic Data Training: Utilizes a synthetic dataset, created with advanced prompts and multiple calls from larger models, to impart new capabilities to SLMs.
- Research Focus: Intended for research purposes to assess SLM abilities and provide a foundation for building more capable frontier models.
- LLaMA-2 Foundation: Inherits the core architecture and general capabilities of its LLaMA-2 base, while focusing on improving reasoning.
Important Considerations
- Research Model Only: Not optimized for general chat or production applications and has not undergone RLHF or DPO training.
- Limitations: Shares limitations with its LLaMA-2 base, including potential data biases, lack of deep contextual understanding, and susceptibility to hallucination.
- System Message Sensitivity: Performance can vary based on system instructions, and it may not show the same few-shot learning gains as larger models.
Orca 2's weights are publicly available to support further research into the development, evaluation, and alignment of SLMs, particularly in the domain of reasoning.