Skip to main content

Generation Parameters

Adjust these settings to control model output.

Key Parameters

Temperature

Controls randomness in responses.

Max Tokens

Maximum length of the response.
Longer max tokens = longer generation time.

Top-p (Nucleus Sampling)

Limits token selection to a cumulative probability.
  • 0.95 (UI default) - Consider tokens until 95% probability mass
  • 0.9 - Slightly more focused
  • 0.5 - Very focused

Top-k

Limits to the k most likely tokens.
  • 50 (default) - Consider top 50 tokens
  • 10 - Very focused
  • 100 - More variety

Parameter Combinations

Factual Q&A

Consistent, accurate answers.

Creative Writing

Varied, creative output.

Code Generation

Precise, syntactically correct code.

Conversation

Natural, varied responses.

Finding the Right Settings

Start with Defaults

Default settings work for most cases:
  • temperature: 0.7
  • max_tokens: 256
  • top_p: 0.95
  • top_k: 50
  • do_sample: true

UI Slider Ranges

The chat interface provides these parameter ranges:

Adjust One at a Time

  1. If responses are too random → lower temperature
  2. If responses are too repetitive → raise temperature
  3. If responses are cut off → increase max_tokens
  4. If responses are too long → decrease max_tokens

Test Systematically

For important applications:
  1. Pick 5-10 test prompts
  2. Try each parameter setting
  3. Compare outputs
  4. Document what works

Advanced Parameters

Repetition Penalty

Reduces repeated phrases.
  • 1.0 - No penalty
  • 1.1 - Mild penalty (recommended)
  • 1.3+ - Strong penalty

Stop Sequences

End generation when these tokens appear.
  • Useful for structured output
  • Example: ["\n\n", "User:"]

Do Sample

Controls whether to use sampling or greedy decoding.
  • true (default) - Use sampling with temperature/top-p/top-k
  • false - Greedy decoding (always pick most likely token)

System Prompt

Set a system message to guide model behavior. Available in the chat interface settings panel. Example system prompts:
  • “You are a helpful coding assistant. Provide concise code examples.”
  • “You are a creative writing partner. Be imaginative and descriptive.”
  • “You are a technical documentation expert. Be precise and thorough.”
The system prompt is prepended to the conversation context and influences how the model responds throughout the session.

Parameter Effects Summary

Next Steps

CLI Training

Train models with CLI

Python API

Programmatic control