Text-to-Text

Overview

Text-to-Text models leverage advanced language models (e.g., Llama3) to transform one piece of text into another—summarizations, translations, or custom language tasks.

Advanced configurations

In both model inference and workflow generative models, you’ll have the opportunity to fine-tune their behavior using various parameters. Adjusting these settings helps strike the right balance between creativity, consistency, and relevance for your specific use case.

Parameters:

• Temperature:
  Controls how deterministic or "creative" the model's responses are.
• Top P (Nucleus Sampling):
  A probability threshold (0 to 1) that limits the model's word choices to those whose cumulative probability does not exceed the specified value.
• Presence Penalty:
  Coefficient that encourages or discourages the model from using words or topics it has already mentioned in the current context.
• Frequency Penalty:
  Similar to presence penalty, but this specifically penalizes words that appear frequently in the output to reduce repetition.
• Max Tokens:
  The maximum number of tokens (roughly approximates to pieces of words) the model will generate in its response.