1.11.3
Diversity-Promoting Methods
What is it?
Diversity-promoting methods are techniques used to reduce repetitive outputs and narrow training behaviour. They operate through sampling controls, repetition penalties and better training-data preparation.
Key points
Temperature scaling flattens the probability distribution when temperature increases. Top-p and Min-p use adaptive probability thresholds. Frequency penalties discourage repeated tokens. Data deduplication removes near-duplicate examples. Multi-domain mixing improves coverage across different tasks. Verbalized sampling asks a model to explicitly express a probability distribution over possible responses.
Combining Diversity Methods
A response-generation system can use temperature scaling and top-p to produce varied candidates. Frequency penalties can reduce repeated wording. During training, deduplication and multi-domain data mixing can reduce over-reliance on repeated patterns and improve domain coverage.
Real-world application
Combine generation-time controls with training-data preparation when building systems that need varied but useful responses. Evaluate both response quality and consistency.
Key takeaway
Diversity is supported by both generation strategies and the quality and variety of training data.