repetition penalty

Repetition penalty decreases probability of previously generated tokens to prevent repetitive output. **Mechanism**: For each token already in output, divide its logit by penalty factor (1.0 = no effect, >1.0 = suppress). Some implementations use additive penalty instead. **Formula**: logit_new = logit / penalty if token appeared, else logit unchanged. **Scope options**: Penalize all previous tokens, sliding window of recent tokens only, or frequency-based (penalize more for repeated tokens). **Typical values**: 1.0-1.2 (subtle), 1.2-1.5 (moderate), 1.5+ (aggressive). **Related techniques**: Presence penalty (flat penalty for any appearance), frequency penalty (scales with occurrence count), no-repeat-ngram (forbid exact n-gram repeats). **Implementation**: Applied before softmax during token selection. **Trade-offs**: Too low → repetitive "loop" outputs, too high → unnatural topic changes, forced vocabulary diversity. **Use cases**: Open-ended generation, chatbots, creative writing. **Best practices**: Start with 1.1-1.2, adjust based on output quality, combine with nucleus sampling for best results.