What is Top-p?

With top_p = 0.9, the model ignores the long tail of unlikely tokens. Providers generally recommend tuning temperature or top_p, not both.