Stack Explorer

Microsoft Orca

llm-model model

Small models that learn from large models

Official site

Supported languages

Orca is a family of Microsoft Research models that learn through imitation of larger models like GPT-4. They use explanation tuning techniques to transfer reasoning capabilities to smaller, more efficient models.

Concepts

explanation-tuningsynthetic-dataknowledge-distillationreasoning-traceschain-of-thought

Pros and Cons

Ventajas

  • + Excellent reasoning for its size
  • + Learns from GPT-4 thought traces
  • + Multiple sizes available
  • + Backed by Microsoft Research
  • + Innovative training techniques
  • + Good benchmark performance

Desventajas

  • - Dependent on GPT-4 synthetic data
  • - Limited by teacher's knowledge
  • - Less versatile than general models
  • - Variable weight availability

Casos de Uso

  • Efficient logical reasoning
  • Step-by-step problem solving
  • Analysis assistants
  • Math tasks
  • Concept explanation