Implementation:Alibaba ROLL SFTConfig

Knowledge Sources	Alibaba ROLL
Domains	Supervised_Learning, Configuration
Last Updated	2026-02-07 20:00 GMT

Overview

Concrete SFT configuration dataclass provided by the Alibaba ROLL library.

Description

The SFTConfig class extends BaseConfig with SFT-specific parameters including dataset field mappings, training hyperparameters, and a single worker configuration for the sft_train cluster.

Usage

Loaded from YAML via Hydra for SFT pipelines.

Code Reference

Source Location

• Repository: Alibaba ROLL
• File: roll/pipeline/sft/sft_config.py
• Lines: L9-64

Signature

@dataclass
class SFTConfig(BaseConfig):
    """
    Configuration for SFT training.

    Attributes:
        pretrain: str - path to pretrained model
        prompt_key: str = "instruction" - dataset prompt key
        response_key: str = "output" - dataset response key
        system_key: Optional[str] - system prompt key
        query_key: Optional[str] - query key
        sft_train: WorkerConfig - SFT worker configuration
        max_grad_norm: float = 1.0 - gradient clipping
    """

Import

from roll.pipeline.sft.sft_config import SFTConfig

I/O Contract

Inputs

Name	Type	Required	Description
YAML config file	str	Yes	Hydra-managed YAML configuration

Outputs

Name	Type	Description
SFTConfig instance	SFTConfig	Validated config with sft_train WorkerConfig

Usage Examples

from hydra import compose, initialize
import dacite
from omegaconf import OmegaConf

initialize(config_path="examples/qwen2.5-7B-sft_megatron")
cfg = compose(config_name="sft_config")
config = dacite.from_dict(data_class=SFTConfig, data=OmegaConf.to_container(cfg, resolve=True))

Related Pages

Implements Principle

• Principle:Alibaba_ROLL_SFT_Configuration

Environment Dependencies

This implementation requires the following environment constraints:

• Environment:Alibaba_ROLL_Python_Runtime_Environment

Heuristics Applied

No specific heuristics apply to this implementation.