00176 使用 trl 进行 sft

前言

此脚本展示了如何使用SFTTrainer将模型或适配器微调到目标数据集中。

src link: https://github.com/huggingface/trl/blob/main/examples/scripts/sft.py

Operating System: Ubuntu 22.04.4 LTS

参考文档

1. TRL - Transformer Reinforcement Learning
2. TRL - Examples (huggingface)
3. examples/scripts/sft.py

训练脚本

"""
# regular:
python examples/scripts/sft.py \
    --model_name_or_path="facebook/opt-350m" \
    --dataset_text_field="text" \
    --report_to="wandb" \
    --learning_rate=1.41e-5 \
    --per_device_train_batch_size=64 \
    --gradient_accumulation_steps=16 \
    --output_dir="sft_openassistant-guanaco" \
    --logging_steps=1 \
    --num_train_epochs=3 \
    --max_steps=-1 \
    --push_to_hub \
    --gradient_checkpointing

# peft:
python examples/scripts/sft.py \
    --model_name_or_path="facebook/opt-350m" \
    --dataset_text_field="text" \
    --report_to="wandb" \
    --learning_rate=1.41e-5 \
    --per_device_train_batch_size=64 \
    --gradient_accumulation_steps=16 \
    --output_dir="sft_openassistant-guanaco" \
    --logging_steps=1 \
    --num_train_epochs=3 \
    --max_steps=-1 \
    --push_to_hub \
    --gradient_checkpointing \
    --use_peft \
    --lora_r=64 \
    --lora_alpha=16
"""

from datasets import load_dataset
from transformers import AutoTokenizer

from trl import (
    ModelConfig,
    SFTConfig,
    SFTScriptArguments,
    SFTTrainer,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)


if __name__ == "__main__":
    parser = TrlParser((SFTScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_config = parser.parse_args_and_config()

    ################
    # Model init kwargs & Tokenizer
    ################
    quantization_config = get_quantization_config(model_config)
    model_kwargs = dict(
        revision=model_config.model_revision,
        trust_remote_code=model_config.trust_remote_code,
        attn_implementation=model_config.attn_implementation,
        torch_dtype=model_config.torch_dtype,
        use_cache=False if training_args.gradient_checkpointing else True,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
    )
    training_args.model_init_kwargs = model_kwargs
    tokenizer = AutoTokenizer.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, use_fast=True
    )
    tokenizer.pad_token = tokenizer.eos_token

    ################
    # Dataset
    ################
    dataset = load_dataset(script_args.dataset_name)

    ################
    # Training
    ################
    trainer = SFTTrainer(
        model=model_config.model_name_or_path,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split],
        tokenizer=tokenizer,
        peft_config=get_peft_config(model_config),
    )

    trainer.train()

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)

结语

第一百七十六篇博文写完，开心！！！！

今天，也是充满希望的一天。