repetition penalty

**Repetition penalty** is the **decoding constraint that down-weights tokens already used in recent output to reduce loops and redundant phrasing** - it is a common safeguard against text degeneration. **What Is Repetition penalty?** - **Definition**: Token-score adjustment that penalizes previously generated tokens during subsequent steps. - **Mechanism**: Modifies logits based on token recurrence history before final token selection. - **Scope Options**: Penalty can apply to full history or a rolling recent-token window. - **Behavioral Effect**: Discourages repeated words, phrases, and structural loops. **Why Repetition penalty Matters** - **Quality Improvement**: Reduces repetitive artifacts that degrade readability. - **Long-Form Stability**: Helps maintain novelty in extended generations. - **User Trust**: Less repetition improves perceived intelligence and polish. - **Operational Robustness**: Prevents worst-case loop failures in interactive systems. - **Decoding Synergy**: Pairs well with temperature and sampling filters for balanced output. **How It Is Used in Practice** - **Penalty Calibration**: Tune magnitude to remove loops without harming essential term reuse. - **Window Design**: Use recent-window penalties for context-sensitive control. - **Domain Testing**: Validate on tasks requiring technical terminology repetition to avoid over-penalization. Repetition penalty is **a key anti-degeneration control in modern decoding stacks** - proper repetition penalties improve readability while preserving semantic accuracy.