Implementation:Huggingface Open r1 TrlParser Usage

Field	Value
Sources	Repo (https://github.com/huggingface/open-r1), Doc (https://huggingface.co/docs/trl)
Domains	NLP, Infrastructure
Last Updated	2026-02-08 00:00 GMT

Overview

Wrapper for TRL's TrlParser that parses CLI arguments and YAML configs into Open-R1's extended configuration dataclasses.

Description

This is a Wrapper Doc. Open-R1 defines custom config dataclasses extending TRL's base classes:

• ScriptArguments — adds dataset_mixture support.
• GRPOScriptArguments — adds reward_funcs, code_provider, cosine parameters, and repetition parameters.
• SFTConfig — adds benchmarks, callbacks, chat_template, and hub_model_revision.
• GRPOConfig — adds num_completions_to_print, system_prompt, and wandb settings.

TrlParser is invoked with a tuple of these dataclass types, and parse_args_and_config() returns a tuple of instances.

Usage

Use at the __main__ entry point of sft.py, grpo.py, or compute_pass_rate.py.

Code Reference

Source

Repository	File	Lines	Description
open-r1	src/open_r1/sft.py	L167-168	SFT usage
open-r1	src/open_r1/grpo.py	L179-180	GRPO usage
open-r1	src/open_r1/configs.py	L22-331	Config dataclasses

Signature

# SFT usage:
parser = TrlParser((ScriptArguments, SFTConfig, ModelConfig))
script_args, training_args, model_args = parser.parse_args_and_config()

# GRPO usage:
parser = TrlParser((GRPOScriptArguments, GRPOConfig, ModelConfig))
script_args, training_args, model_args = parser.parse_args_and_config()

Import

from trl import TrlParser, ModelConfig
from open_r1.configs import ScriptArguments, SFTConfig, GRPOConfig, GRPOScriptArguments

I/O Contract

Inputs

Input	Type	Description
CLI arguments or YAML config file path	sys.argv / file path	Raw command-line arguments or a --config flag pointing to a YAML file
Dataclass types tuple	tuple[type, ...]	Tuple of dataclass types to parse into (e.g., (ScriptArguments, SFTConfig, ModelConfig))

Outputs

Output	Type	Description
Config tuple	tuple[ScriptArguments, TrainingConfig, ModelConfig]	Validated, typed configuration objects controlling all aspects of the training run

Usage Examples

Invoking via CLI with accelerate:

accelerate launch src/open_r1/sft.py --config recipes/OpenR1-Distill-7B/sft/config_distill.yaml

This command launches the SFT training script. TrlParser reads the YAML config file specified by --config, merges it with dataclass defaults and any additional CLI flags, validates the result, and returns the typed config tuple to the training function.

Related Pages

Principle:Huggingface_Open_r1_Configuration_Parsing