LLM Sampling Parameters Playground: Temperature, Top-k & Top-p

This playground allows you to move beyond simply accepting an LLM's output and instead, understand the mathematical mechanics behind it. At its core, we simulate how different sampling parameters manipulate the raw probability distribution of possible next tokens.

You can observe three key levers: Temperature (controlling randomness), Top-k (limiting choices to the top K tokens), and Top-p (cumulative probability mass). By adjusting these sliders, you watch the probability chart reshape in real-time. The optional model inspection feature provides deep visibility into actual logprobs, ensuring that all calculations remain 100% private within your browser.

We recommend a methodical approach when optimizing your sampling parameters. Start by defining the desired output persona (e.g., 'Academic,' 'Casual Chatbot,' or 'Coder').

• For Technical/Factual Tasks: Set Temperature low (0.2 - 0.5). Use Top-p around 0.9 to keep the focus on high-probability tokens while maintaining slight flexibility.
• For Creative/Narrative Tasks: Increase Temperature (0.7 - 1.0) and use a moderate Top-p (0.8 - 0.95) to allow for unexpected, yet relevant, vocabulary choices.
• Debugging Coherence Issues: If the output suddenly becomes repetitive or stuck in loops, try slightly decreasing your temperature first; this often stabilizes the model's focus.