|
| 1 | +# Copyright 2020-2025 The HuggingFace Team. All rights reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +from dataclasses import dataclass, field |
| 16 | +from typing import Any |
| 17 | + |
| 18 | +from transformers import TrainingArguments |
| 19 | + |
| 20 | + |
| 21 | +@dataclass |
| 22 | +class ORPOConfig(TrainingArguments): |
| 23 | + r""" |
| 24 | + Configuration class for the [`experimental.orpo.ORPOTrainer`]. |
| 25 | +
|
| 26 | + This class includes only the parameters that are specific to ORPO training. For a full list of training arguments, |
| 27 | + please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may |
| 28 | + differ from those in [`~transformers.TrainingArguments`]. |
| 29 | +
|
| 30 | + Using [`~transformers.HfArgumentParser`] we can turn this class into |
| 31 | + [argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the |
| 32 | + command line. |
| 33 | +
|
| 34 | + Parameters: |
| 35 | + max_length (`int` or `None`, *optional*, defaults to `1024`): |
| 36 | + Maximum length of the sequences (prompt + completion) in the batch. This argument is required if you want |
| 37 | + to use the default data collator. |
| 38 | + max_prompt_length (`int` or `None`, *optional*, defaults to `512`): |
| 39 | + Maximum length of the prompt. This argument is required if you want to use the default data collator. |
| 40 | + max_completion_length (`int`, *optional*): |
| 41 | + Maximum length of the completion. This argument is required if you want to use the default data collator |
| 42 | + and your model is an encoder-decoder. |
| 43 | + beta (`float`, *optional*, defaults to `0.1`): |
| 44 | + Parameter controlling the relative ratio loss weight in the ORPO loss. In the |
| 45 | + [paper](https://huggingface.co/papers/2403.07691), it is denoted by λ. In the |
| 46 | + [code](https://github.com/xfactlab/orpo), it is denoted by `alpha`. |
| 47 | + disable_dropout (`bool`, *optional*, defaults to `True`): |
| 48 | + Whether to disable dropout in the model. |
| 49 | + label_pad_token_id (`int`, *optional*, defaults to `-100`): |
| 50 | + Label pad token id. This argument is required if you want to use the default data collator. |
| 51 | + padding_value (`int`, *optional*): |
| 52 | + Padding value to use. If `None`, the padding value of the tokenizer is used. |
| 53 | + truncation_mode (`str`, *optional*, defaults to `"keep_end"`): |
| 54 | + Truncation mode to use when the prompt is too long. Possible values are `"keep_end"` or `"keep_start"`. |
| 55 | + This argument is required if you want to use the default data collator. |
| 56 | + generate_during_eval (`bool`, *optional*, defaults to `False`): |
| 57 | + If `True`, generates and logs completions from the model to W&B or Comet during evaluation. |
| 58 | + is_encoder_decoder (`bool`, *optional*): |
| 59 | + When using the `model_init` argument (callable) to instantiate the model instead of the `model` argument, |
| 60 | + you need to specify if the model returned by the callable is an encoder-decoder model. |
| 61 | + model_init_kwargs (`dict[str, Any]`, *optional*): |
| 62 | + Keyword arguments to pass to `AutoModelForCausalLM.from_pretrained` when instantiating the model from a |
| 63 | + string. |
| 64 | + dataset_num_proc (`int`, *optional*): |
| 65 | + Number of processes to use for processing the dataset. |
| 66 | + """ |
| 67 | + |
| 68 | + _VALID_DICT_FIELDS = TrainingArguments._VALID_DICT_FIELDS + ["model_init_kwargs"] |
| 69 | + |
| 70 | + # Parameters whose default values are overridden from TrainingArguments |
| 71 | + learning_rate: float = field( |
| 72 | + default=1e-6, |
| 73 | + metadata={"help": "The initial learning rate for AdamW."}, |
| 74 | + ) |
| 75 | + logging_steps: float = field( |
| 76 | + default=10, |
| 77 | + metadata={ |
| 78 | + "help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, " |
| 79 | + "will be interpreted as ratio of total training steps." |
| 80 | + }, |
| 81 | + ) |
| 82 | + gradient_checkpointing: bool = field( |
| 83 | + default=True, |
| 84 | + metadata={ |
| 85 | + "help": "If True, use gradient checkpointing to save memory at the expense of slower backward pass." |
| 86 | + }, |
| 87 | + ) |
| 88 | + bf16: bool | None = field( |
| 89 | + default=None, |
| 90 | + metadata={ |
| 91 | + "help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA " |
| 92 | + "architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if " |
| 93 | + "`fp16` is not set." |
| 94 | + }, |
| 95 | + ) |
| 96 | + # Transformers 4.57.0 introduced a bug that caused the dtype of `lr_scheduler_kwargs` to be unparsable. This issue |
| 97 | + # was fixed in https://github.com/huggingface/transformers/pull/41322, but the fix has not yet been released. We |
| 98 | + # add a temporary workaround here, which can be removed once the fix is available—likely in Transformers 4.57.2. |
| 99 | + lr_scheduler_kwargs: dict | str | None = field( |
| 100 | + default=None, |
| 101 | + metadata={ |
| 102 | + "help": "Additional parameters for the lr_scheduler, such as {'num_cycles': 1} for cosine with hard " |
| 103 | + "restarts." |
| 104 | + }, |
| 105 | + ) |
| 106 | + |
| 107 | + max_length: int | None = field( |
| 108 | + default=1024, |
| 109 | + metadata={"help": "Maximum length of the sequences (prompt + completion) in the batch."}, |
| 110 | + ) |
| 111 | + max_prompt_length: int | None = field( |
| 112 | + default=512, |
| 113 | + metadata={ |
| 114 | + "help": "Maximum length of the prompt. This argument is required if you want to use the default data " |
| 115 | + "collator and your model is an encoder-decoder." |
| 116 | + }, |
| 117 | + ) |
| 118 | + max_completion_length: int | None = field( |
| 119 | + default=None, |
| 120 | + metadata={ |
| 121 | + "help": "Maximum length of the completion. This argument is required if you want to use the default data " |
| 122 | + "collator and your model is an encoder-decoder." |
| 123 | + }, |
| 124 | + ) |
| 125 | + beta: float = field( |
| 126 | + default=0.1, |
| 127 | + metadata={ |
| 128 | + "help": "Parameter controlling the relative ratio loss weight in the ORPO loss. In the paper, it is " |
| 129 | + "denoted by λ." |
| 130 | + }, |
| 131 | + ) |
| 132 | + disable_dropout: bool = field( |
| 133 | + default=True, |
| 134 | + metadata={"help": "Whether to disable dropout in the model."}, |
| 135 | + ) |
| 136 | + label_pad_token_id: int = field( |
| 137 | + default=-100, |
| 138 | + metadata={ |
| 139 | + "help": "Label pad token id. This argument is required if you want to use the default data collator." |
| 140 | + }, |
| 141 | + ) |
| 142 | + padding_value: int | None = field( |
| 143 | + default=None, |
| 144 | + metadata={"help": "Padding value to use. If `None`, the padding value of the tokenizer is used."}, |
| 145 | + ) |
| 146 | + truncation_mode: str = field( |
| 147 | + default="keep_end", |
| 148 | + metadata={ |
| 149 | + "help": "Truncation mode to use when the prompt is too long.", |
| 150 | + "choices": ["keep_end", "keep_start"], |
| 151 | + }, |
| 152 | + ) |
| 153 | + generate_during_eval: bool = field( |
| 154 | + default=False, |
| 155 | + metadata={"help": "If `True`, generates and logs completions from the model to W&B during evaluation."}, |
| 156 | + ) |
| 157 | + is_encoder_decoder: bool | None = field( |
| 158 | + default=None, |
| 159 | + metadata={ |
| 160 | + "help": "When using the `model_init` argument (callable) to instantiate the model instead of the `model` " |
| 161 | + "argument, you need to specify if the model returned by the callable is an encoder-decoder model." |
| 162 | + }, |
| 163 | + ) |
| 164 | + model_init_kwargs: dict[str, Any] | None = field( |
| 165 | + default=None, |
| 166 | + metadata={ |
| 167 | + "help": "Keyword arguments to pass to `AutoModelForCausalLM.from_pretrained` when instantiating the model " |
| 168 | + "from a string." |
| 169 | + }, |
| 170 | + ) |
| 171 | + dataset_num_proc: int | None = field( |
| 172 | + default=None, |
| 173 | + metadata={"help": "Number of processes to use for processing the dataset."}, |
| 174 | + ) |
| 175 | + |
| 176 | + def __post_init__(self): |
| 177 | + self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16 |
| 178 | + |
| 179 | + super().__post_init__() |
0 commit comments