If an AI keeps circling back to the same words, frequency penalty is the nudge that says, “We’ve already used that one.” It reduces repetitive or redundant output by making previously sampled tokens less likely. section 2.10
Use it as a creativity-versus-coherence dial:
- The model adjusts its token probabilities after words or token sequences appear in the output. section 2.10
- Higher values discourage repetition more strongly; lower values leave the model more flexibility. The described range is −2.0 to 2.0. section 2.10
- Unlike presence penalty, frequency penalty responds to how often a token has appeared, rather than applying a one-time penalty after its first appearance. section 2.11
From ordinary smartphone ideas to unusual ones
Without a frequency penalty, a model may suggest familiar features such as a high-resolution display, multiple cameras, fast processing, long battery life, and a sleek design. Increasing the penalty can steer it toward less common ideas such as a holographic display, graphene-based battery, modular hardware, real-time translation, or palm-vein authentication. Check that the result remains useful; novelty alone is not a quality guarantee. section 2.10
|
|
Generate some ideas for a new smartphone. |
|
|
Sure! Here are a few ideas: a high-resolution display, multiple camera lenses, fast processing speed, long battery life, and a sleek design. |
Compare the ordinary smartphone suggestions with the higher-novelty alternative in the paired output tables. section 2.10
