single source · post-training
SFT followed by reinforcement learning and on-policy distillation
Uses supervised fine-tuning, then reinforcement learning, followed by on-policy distillation as a standard post-training sequence.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See post-training for the methods filed beside it.
Filed under post-training
Related methods
Supported by
Methods the sources say are used for this one.
We use large-scale asynchronous RL in synthesized tasks to improve model performance
we invest in large-scale, automated pipelines for data synthesis and environment construction
DeepSeek V41 Tech Report · implied- Checkpoint merging for successive RL runsconfigures it
we use model merging to reinitialize successive RL runs
- Dynamic Post-Training Recipe Reconfigurationconfigures it
We continuously track model capabilities and may adjust the training recipe accordingly
and 5 more
- Failure replay and targeted reinforcement learningapplied to it
the pipeline enables systematic replay of failures and targeted reinforcement learning against observed model weaknesses
the final full-vocabulary OPD task is trained on datasets from all domains using over 40 teacher models
Whenever a task is used in a new RL run, the resulting trajectories provide fresh evidence for quality re-auditing
DeepSeek V41 Tech Report · impliedprocedurally constructs and scales interactive agent environments in which trajectories can be collected and evaluated at low cost
synthesizes diverse, verifiable training tasks together with their reference solutions and reward signals
DeepSeek V41 Tech Report · implied
Filed alongside
Other methods under post-training.