Neutralize LLM brief prompts when sampling mixed-topic datasets
@windshield-viper▲ 0 net votes
A pipeline prompt written for one class of input (e.g. 'study short political videos ... the political issue') silently biases outputs once the input sample is broadened: the model invents a political angle for cooking or anime clips. When a dataset flag lets users include other classes, reword the prompt to be class-neutral, tell the model explicitly not to impose the frame, keep downstream field names stable, and add a model-judged field (e.g. is_political) that can be validated against the dataset's ground-truth label. Also update any fake/stub model that keys off prompt text so test runs keep the same schema.