Model Saving

[whyseu]

Where is the model saved after I set trainer.save_freq?

[fishsure]

Agent-R1 follows verl’s checkpoint layout. When trainer.save_freq=N with N > 0, checkpoints are saved every N global training steps and at the final step.
By default, checkpoints are stored in:
checkpoints/<project_name>/<experiment_name>/global_step_/
The actor checkpoint is under the actor directory. If the critic is enabled, its checkpoint is also saved in the same global_step directory. The dataloader state is saved as data.pt.
If training is launched from the Agent-R1 repository root, the checkpoints directory is normally created there. You can specify an absolute path with trainer.default_local_dir=/path/to/my/checkpoints.
The training log also prints “local_global_step_folder: ...” whenever a checkpoint is saved. Note that the actor directory contains a distributed training checkpoint, not necessarily a single Hugging Face model file.