Open Models Almanac
Methodspost-trainingrollout & RL infrastructure

specific method · filed under post-training

Partial rollout

It partitions a trajectory that is too long to finish in one step across multiple steps, while limiting staleness and the share of partial samples in each batch.1

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. to partition overlong trajectories across steps, while limiting staleness and the proportion of partial samples in each batch — MiMo-V2-Flash Technical Report

Also called Partial rollout for mixed-task RL, partial rollout for synchronous RL, partial rollout scheme, partial rollouts.

sources
3
models
4
labs adopt it
2
strongest
used

Documented in

Evidence

5 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
used 5

MiMo-V2.6-Flash · MiMo-V2.6-Pro

Xiaomiadopted1 source

we use partial rollout to keep the running batch saturated: when we collect a training batch, we leave the in-flight sequences interrupted and resume them in the next rollout phase

usedsoftware implementationin MiMo-V2.6 RL trainingXiaomi
1 more from these models

we have maintained a Data Scheduler that targets a specified training distribution across data sources, together with dynamic sampler and partial rollout

usedunclearin MiMo-V2.6 RL and OPD infrastructureXiaomi

Kimi K3

Moonshot AIadopted1 source

To mitigate the long-tail latency that intensifies in long-horizon tasks, we extend the partial rollout scheme from our synchronous RL framework

usedunclearin Kimi K3Moonshot AI
1 more from this model

and use partial rollouts [118] to reduce tail latency from ultra-long trajectories.

usedpost trainingin Kimi K3Moonshot AI
↑