aurora

1
# Inspired by: https://github.com/CarperAI/trlx/blob/main/examples/summarize_rlhf/reward_model/train_reward_model_gptj.py
2

3
from typing import TYPE_CHECKING, Optional, List
4
from transformers import Seq2SeqTrainingArguments
5

6
from llmtuner.data import get_dataset, preprocess_dataset, split_dataset
7
from llmtuner.extras.callbacks import SavePeftModelCallback
8
from llmtuner.extras.ploting import plot_loss
9
from llmtuner.model import load_model_and_tokenizer
10
from llmtuner.train.rm.collator import PairwiseDataCollatorWithPadding
11
from llmtuner.train.rm.metric import compute_accuracy
12
from llmtuner.train.rm.trainer import PairwiseTrainer
13
from llmtuner.train.utils import create_modelcard_and_push
14

15
if TYPE_CHECKING:
16
    from transformers import TrainerCallback
17
    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments
18

19

20
def run_rm(
21
    model_args: "ModelArguments",
22
    data_args: "DataArguments",
23
    training_args: "Seq2SeqTrainingArguments",
24
    finetuning_args: "FinetuningArguments",
25
    callbacks: Optional[List["TrainerCallback"]] = None
26
):
27
    dataset = get_dataset(model_args, data_args)
28
    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train, add_valuehead=True)
29
    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="rm")
30
    data_collator = PairwiseDataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
31

32
    # Update arguments
33
    training_args_dict = training_args.to_dict()
34
    training_args_dict.update(dict(remove_unused_columns=False)) # important for pairwise dataset
35
    training_args = Seq2SeqTrainingArguments(**training_args_dict)
36

37
    # Initialize our Trainer
38
    trainer = PairwiseTrainer(
39
        model=model,
40
        args=training_args,
41
        tokenizer=tokenizer,
42
        data_collator=data_collator,
43
        callbacks=callbacks + [SavePeftModelCallback()],
44
        compute_metrics=compute_accuracy,
45
        **split_dataset(dataset, data_args, training_args)
46
    )
47

48
    # Training
49
    if training_args.do_train:
50
        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
51
        trainer.save_model()
52
        trainer.log_metrics("train", train_result.metrics)
53
        trainer.save_metrics("train", train_result.metrics)
54
        trainer.save_state()
55
        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
56
            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
57

58
    # Evaluation
59
    if training_args.do_eval:
60
        metrics = trainer.evaluate(metric_key_prefix="eval")
61
        trainer.log_metrics("eval", metrics)
62
        trainer.save_metrics("eval", metrics)
63

64
    # Predict
65
    if training_args.do_predict:
66
        predict_results = trainer.predict(dataset, metric_key_prefix="predict")
67
        trainer.log_metrics("predict", predict_results.metrics)
68
        trainer.save_metrics("predict", predict_results.metrics)
69
        trainer.save_predictions(predict_results)
70

71
    # Create model card
72
    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)
73