Implementation:Huggingface Open r1 Get Callbacks

Template:Metadata

Overview

Concrete tool for resolving callback names to TrainerCallback instances from Open-R1's callback registry.

Description

The get_callbacks function resolves callback names from config to TrainerCallback instances. The CALLBACKS registry currently contains one entry: "push_to_hub_revision" -> PushToHubRevisionCallback. PushToHubRevisionCallback implements on_save to push checkpoints to unique Hub branches and optionally trigger benchmark evaluation via Slurm. The callback uses a DummyConfig workaround to avoid breaking the accelerator distributed state when modifying training arguments.

Usage

Import when initializing SFTTrainer or GRPOTrainer to add configured callbacks.

Code Reference

Source

Repository: open-r1, File: src/open_r1/utils/callbacks.py, Lines: L43-92

Signature

def get_callbacks(train_config, model_config) -> List[TrainerCallback]:
    """Resolve callback names from config to TrainerCallback instances."""

class PushToHubRevisionCallback(TrainerCallback):
    def __init__(self, model_config) -> None:
    def on_save(self, args, state, control, **kwargs):

Import

from open_r1.utils.callbacks import get_callbacks

I/O Contract

Inputs

Name	Type	Required	Description
train_config	SFTConfig or GRPOConfig	Yes	Training configuration with callbacks list of strings
model_config	ModelConfig	Yes	Model configuration with trust_remote_code

Outputs

Type	Description
list[TrainerCallback]	Callback instances registered in training loop

Usage Examples

Configuring callbacks in YAML:

# sft_config.yaml
callbacks:
  - push_to_hub_revision

Resolving callbacks in training script:

from open_r1.utils.callbacks import get_callbacks

callbacks = get_callbacks(train_config, model_config)
trainer = SFTTrainer(
    model=model,
    args=train_config,
    callbacks=callbacks,
    ...
)

Related Pages

• Principle:Huggingface_Open_r1_Training_Callbacks