single mention · post-training
Sample Mixer
Sample Mixer is a sample-mixing mechanism for stable asynchronous mixed-task RL that preserves a specified training distribution across data sources.1 It combines four mechanisms: adaptive rollout concurrency sets per-source budgets, adaptive scheduling selects sources within budgets, predictive dispatch places rollouts across ranks, and sample replay covers startup and recovery.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Mixed-task RL must preserve this distribution de- spite large variations in rollout duration and filtering rates, so that every data source is trained effectively
— MiMo V2 6 technical reportAdaptive Rollout Concurrency sets per-source budgets, Adaptive Rollout Scheduling selects data sources within the budgets,Predictive Rollout Dispatchplaces new rollouts across ranks, andSample Replay covers startup and recovery
— MiMo V2 6 technical report
Also called A sample mixing mechanism.
Filed under post-training › rollout & RL infrastructure