aurora

1
# Inspired by: https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.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.constants import IGNORE_INDEX
8
from llmtuner.extras.ploting import plot_loss
9
from llmtuner.hparams import ModelArguments
10
from llmtuner.model import load_model_and_tokenizer
11
from llmtuner.train.dpo.collator import DPODataCollatorWithPadding
12
from llmtuner.train.dpo.trainer import CustomDPOTrainer
13
from llmtuner.train.utils import create_modelcard_and_push, create_ref_model
14

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

19

20
def run_dpo(
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)
29
    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="rm")
30
    data_collator = DPODataCollatorWithPadding(
31
        tokenizer=tokenizer,
32
        pad_to_multiple_of=8,
33
        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
34
    )
35

36
    # Create reference model
37
    if finetuning_args.ref_model is None and (not training_args.do_train): # use the model itself
38
        ref_model = model
39
    else:
40
        ref_model = create_ref_model(model_args, finetuning_args)
41

42
    # Update arguments
43
    training_args_dict = training_args.to_dict()
44
    training_args_dict.update(dict(remove_unused_columns=False)) # important for pairwise dataset
45
    training_args = Seq2SeqTrainingArguments(**training_args_dict)
46

47
    # Initialize our Trainer
48
    trainer = CustomDPOTrainer(
49
        beta=finetuning_args.dpo_beta,
50
        model=model,
51
        ref_model=ref_model,
52
        args=training_args,
53
        tokenizer=tokenizer,
54
        data_collator=data_collator,
55
        callbacks=callbacks,
56
        **split_dataset(dataset, data_args, training_args)
57
    )
58

59
    # Training
60
    if training_args.do_train:
61
        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
62
        trainer.save_model()
63
        trainer.log_metrics("train", train_result.metrics)
64
        trainer.save_metrics("train", train_result.metrics)
65
        trainer.save_state()
66
        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
67
            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
68

69
    # Evaluation
70
    if training_args.do_eval:
71
        metrics = trainer.evaluate(metric_key_prefix="eval")
72
        if id(model) == id(ref_model): # unable to compute rewards without a reference model
73
            remove_keys = [key for key in metrics.keys() if "rewards" in key]
74
            for key in remove_keys:
75
                metrics.pop(key)
76
        trainer.log_metrics("eval", metrics)
77
        trainer.save_metrics("eval", metrics)
78

79
    # Create model card
80
    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)
81