single source · other
Re-render rollout context into training segments
Re-rendering splits a rollout sequence into segments whenever tokenization or chat-template changes alter the previously rendered context, instead of appending new tokens to the unchanged token-in/token-out context.1 Training then computes losses only on tokens the model newly generates in each segment, masking the re-rendered context and system, user and tool-response tokens.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
We split sequences into segments due to, e.g., chat-template changes that adapt previously rendered contexts
— tech reportWe only train on the tokens that are newly generated by the model in each segment, masking out re-rendered context as well as system, user and tool-response tokens
— tech report
Also called Segments.
Filed under other