The Cohere Playground allows for the exploration of different endpoints (
Likelihood), model sizes (
orca), and settings (temperature, etc.). Begin experimentation using
orca, the most capable model. Once a specific use-case is identified, swap out
orca with different baseline models to identify nuances in output and pick the most appropriate model.
New language model users should begin with
Generate. Try out different tasks like summarization, Q&A, creative writing, or abstract reasoning to better understand capabilities. After using
Generate, begin exploring other endpoints using the same iterative process.
Prompt structure is directly impactful to response quality. Review documentation on prompt engineering to maximize model performance.
Models of larger sizes are more capable of complex tasks but models of smaller sizes have faster response times. Here is a rough guideline for which model size to use for various tasks:
orca(most capable model): abstract reasoning, Q&A, question classification
shark: data extraction, copy generation
seal: summarization, content rephrasing, simple Q&A
otter: semantic relatedness, creative writing
shrimp(fastest model): sentiment analysis, simple classification
The models are listed above from largest to smallest. Larger models can perform all tasks accomplished by smaller models.