Google

A repetition penalty is a mechanism designed to discourage a model from generating repetitive or redundant text. It functions by measuring the frequency of repeated tokens or phrases within the generated sequence and applying a penalty that is proportional to their occurrence, thereby preventing monotonous output.

Repetition Penalty

A representation-based repetition penalty is a technique to discourage repetitive text by operating on the model's internal representations. This method calculates a penalty based on the maximum distance between the representation of the token being predicted and the representations of tokens that have already been generated. By penalizing tokens whose representations are too close to previous ones, the search objective is guided away from producing degenerate, repetitive outputs.

Representation-based Repetition Penalty

Based on the case study, analyze the trade-off demonstrated by the developer's adjustment. Explain why the second output, despite being less repetitive, is of lower quality and what this reveals about the effect of an overly aggressive penalty.

Tuning Text Generation for Poetry

A developer is using a text generation model to summarize articles. They notice the summaries often contain repetitive phrases like 'The article discusses... The article also discusses...'. To fix this, they apply a strong penalty to tokens that have already appeared in the generated text. Describe two potential negative consequences of setting this penalty value *excessively high*.

Consequences of Over-Tuning a Repetition Penalty

A language model is tasked with writing a short story and produces the following output: 'The knight rode his horse through the dark forest. The forest was very dark. The knight was a brave knight, and the dark forest did not scare him.' Which of the following adjustments to the generation process would be most effective at discouraging this kind of repetitive phrasing?

Learn Before

Related