hard20 pts · first answer countsllm-fundamentalsllm-basics

Training on its own outputs

A team repeatedly trains new model versions mostly on text generated by earlier versions, with little fresh human data. What risk grows over generations?

A.The tokenizer's vocabulary doubles in size each timeB.Model collapse, where rare cases fade and outputs grow sameyC.The context window shrinks by half each generationD.The models stop being able to produce any text at all

Discussion 0

Join free to post in the discussion.

Training on its own outputs

Discussion 0

Related questions