Back to Home
Inside an LLM
Top-P Sampling
Top-P controls how many likely next words the AI is allowed to choose from.
This is a simple learning demo using predefined word probabilities, not a real AI model.
Interactive Playground
0.90
0.1
Narrow 0.5
Moderate 1.0
All words
Narrow 0.5
Moderate 1.0
All words
Sentence So Far
Your sentence will grow here…
Last chosen:
Words added: 0
Live Visualization
Cumulative Probability
0%
Top-P cutoff: 90%
100%
Word Probabilities
InAllowed
OutExcluded
Statistics
0.90
Top-P Value
Cum. Probability
Allowed Words
Excluded Words
Chosen Word
How It Works
📊
Predicted
Words
Words
→
📉
Sort by
Prob.
Prob.
→
✂️
Keep
Top-P
Top-P
→
🎯
Choose
Word
Word
🎓
Beginner Tips
2 tips
P=1.0 considers all words (maximum randomness). P=0.1 restricts to only the very top few words, near-deterministic output.
Top-P and Temperature both control randomness but differently. Most apps use them together (e.g. temperature=0.7, top_p=0.9).
💡
Key Takeaway
Top-P limits the AI to a group of likely words, helping control randomness while still allowing variety.
A low Top-P (e.g. 0.3) means only the top 1-2 words are available. A high Top-P (e.g. 0.95) opens up many more options. Most AI systems use 0.9 by default.