File size: 5,166 Bytes
2f5127c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
# Copyright 2020-2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import importlib
import os
import sys
from dataclasses import dataclass, field
from typing import Optional

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer

from trl import GRPOConfig, GRPOTrainer, ModelConfig, ScriptArguments, TrlParser, get_peft_config
from trl.rewards import think_format_reward


reward_funcs_registry = {
    "think_format_reward": think_format_reward,
}


@dataclass
class GRPOScriptArguments(ScriptArguments):
    """
    Script arguments for the GRPO training script.

    Args:
        reward_model_name_or_path (`str` or `None`, *optional*, defaults to `None`):
            Reward model id of a pretrained model hosted inside a model repo on huggingface.co or local path to a
            directory containing model weights saved using [`~transformers.PreTrainedModel.save_pretrained`].
        reward_funcs (`list[str]` or `None`, *optional*, defaults to `None`):
            Reward functions to use. It can be either one of  `"think_format_reward"`; or a dotted import path "
            (e.g., `'my_lib.rewards.custom_reward'`).
    """

    reward_model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": "Reward model id of a pretrained model hosted inside a model repo on huggingface.co or "
            "local path to a directory containing model weights saved using `PreTrainedModel.save_pretrained`."
        },
    )
    reward_funcs: Optional[list[str]] = field(
        default=None,
        metadata={
            "help": "Reward functions to use. It can be either one of  'think_format_reward'; or a dotted "
            "import path. (e.g., 'my_lib.rewards.custom_reward')."
        },
    )


def main(script_args, training_args, model_args):
    # Load a pretrained model
    model = AutoModelForCausalLM.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )

    # Get the reward models and functions
    reward_funcs = []
    if script_args.reward_model_name_or_path:
        reward_model = AutoModelForSequenceClassification.from_pretrained(
            script_args.reward_model_name_or_path, trust_remote_code=model_args.trust_remote_code, num_labels=1
        )
        reward_funcs.append(reward_model)

    if script_args.reward_funcs:
        for func_name in script_args.reward_funcs:
            if func_name in reward_funcs_registry:
                reward_funcs.append(reward_funcs_registry[func_name])
            elif "." in func_name:
                module_path, func_name = func_name.rsplit(".", 1)
                sys.path.insert(0, os.getcwd())
                module = importlib.import_module(module_path)
                reward_func = getattr(module, func_name)
                reward_funcs.append(reward_func)
            else:
                raise ValueError(
                    f"Could not load reward function '{func_name}'. Expected one of "
                    f"{list(reward_funcs_registry.keys())} or a valid import path."
                )

    # Load the dataset
    dataset = load_dataset(script_args.dataset_name, name=script_args.dataset_config)

    # Initialize the GRPO trainer
    trainer = GRPOTrainer(
        model=model,
        reward_funcs=reward_model,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
        processing_class=tokenizer,
        peft_config=get_peft_config(model_args),
    )

    # Train and push the model to the Hub
    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)


def make_parser(subparsers: argparse._SubParsersAction = None):
    dataclass_types = (GRPOScriptArguments, GRPOConfig, ModelConfig)
    if subparsers is not None:
        parser = subparsers.add_parser("grpo", help="Run the GRPO training script", dataclass_types=dataclass_types)
    else:
        parser = TrlParser(dataclass_types)
    return parser


if __name__ == "__main__":
    parser = make_parser()
    script_args, training_args, model_args = parser.parse_args_and_config()
    main(script_args, training_args, model_args)