h5PSSKJrJr SSKJrJr SSKJr \"SS\55rg)) dataclassfield)AnyOptional)TrainingArgumentsc^\rSrSr%Sr\R S/-r\"SSS0S9r\ \ S'\"S SS 0S9r \ \ S '\"S SS 0S9r \ \ S'\"SSS0S9r\\ \ S'\"SSS0S9r\\\\4\ S'\"SSS0S9r\\\ S'\"SSS0S9r\\ S'\"SSS0S9r\\\\4\ S'\"SSS0S9r\\\ S'\"SSS0S9r\\\ S'\"SSS0S9r\\\ S'\"S SS!0S9r\\\ S"'\"S#SS$0S9r\ \ S%'\"S&SS'0S9r\\ S('\"S#SS)0S9r\ \ S*'\"SSS+0S9r\\\ S,'\"SSS-0S9r \\ \ S.'\"SSS/0S9r!\\ \ S0'\"S#SS10S9r"\ \ S2'\"S3SS40S9r#\\ S5'\"S#SS60S9r$\ \ S7'U4S8jr%S9r&U=r'$): SFTConfiga Configuration class for the [`SFTTrainer`]. This class includes only the parameters that are specific to SFT training. For a full list of training arguments, please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may differ from those in [`~transformers.TrainingArguments`]. Using [`~transformers.HfArgumentParser`] we can turn this class into [argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the command line. Parameters: > Parameters that control the model model_init_kwargs (`dict[str, Any]` or `None`, *optional*, defaults to `None`): Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model` argument of the [`SFTTrainer`] is provided as a string. If you're training a MoE architecture and want to include the load balancing/auxilliary loss as a part of the final loss, remember to set `output_router_logits=True` in this dictionary. chat_template_path (`str` or `None`, *optional*, defaults to `None`): If specified, sets the model's chat template. This can either be the path to a tokenizer (local directory or Hugging Face Hub model) or a direct path to a Jinja template file. When using a Jinja file, you must ensure that any special tokens referenced in the template are added to the tokenizer and that the model's embedding layer is resized accordingly. > Parameters that control the data preprocessing dataset_text_field (`str`, *optional*, defaults to `"text"`): Name of the column that contains text data in the dataset. dataset_kwargs (`dict[str, Any]` or `None`, *optional*, defaults to `None`): Dictionary of optional keyword arguments for the dataset preparation. The only supported key is `skip_prepare_dataset`. When the model is a VLM, `skip_prepare_dataset` is automatically treated as `True` regardless of the provided value, since preprocessing is done on the fly. dataset_num_proc (`int` or `None`, *optional*, defaults to `None`): Number of processes to use for processing the dataset. eos_token (`str` or `None`, *optional*, defaults to `None`): Token used to indicate the end of a turn or sequence. If `None`, it defaults to `processing_class.eos_token`. pad_token (`int` or `None`, *optional*, defaults to `None`): Token used for padding. If `None`, it defaults to `processing_class.pad_token`, or if that is also `None`, it falls back to `processing_class.eos_token`. max_length (`int` or `None`, *optional*, defaults to `1024`): Maximum length of the tokenized sequence. Sequences longer than `max_length` are truncated from the right. If `None`, no truncation is applied. When packing is enabled, this value sets the sequence length. packing (`bool`, *optional*, defaults to `False`): Whether to group multiple sequences into fixed-length blocks to improve computational efficiency and reduce padding. Uses `max_length` to define sequence length. packing_strategy (`str`, *optional*, defaults to `"bfd"`): Strategy for packing sequences. Can be either `"bfd"` (best-fit decreasing, default), or `"wrapped"`. padding_free (`bool`, *optional*, defaults to `False`): Whether to perform forward passes without padding by flattening all sequences in the batch into a single continuous sequence. This reduces memory usage by eliminating padding overhead. Currently, this is only supported with the FlashAttention 2 or 3, which can efficiently handle the flattened batch structure. When packing is enabled with strategy `"bfd"`, padding-free is enabled, regardless of the value of this parameter. pad_to_multiple_of (`int` or `None`, *optional*, defaults to `None`): If set, the sequences will be padded to a multiple of this value. eval_packing (`bool` or `None`, *optional*, defaults to `None`): Whether to pack the eval dataset. If `None`, uses the same value as `packing`. > Parameters that control the training completion_only_loss (`bool` or `None`, *optional*, defaults to `None`): Whether to compute loss only on the completion part of the sequence. If set to `True`, loss is computed only on the completion, which is supported only for [prompt-completion](#prompt-completion) datasets. If `False`, loss is computed on the entire sequence. If `None` (default), the behavior depends on the dataset: loss is computed on the completion for [prompt-completion](#prompt-completion) datasets, and on the full sequence for [language modeling](#language-modeling) datasets. assistant_only_loss (`bool`, *optional*, defaults to `False`): Whether to compute loss only on the assistant part of the sequence. If set to `True`, loss is computed only on the assistant responses, which is supported only for [conversational](#conversational) datasets. If `False`, loss is computed on the entire sequence. loss_type (`str`, *optional*, defaults to `"nll"`): Type of loss to use. Possible values are `"nll"` (negative log-likelihood, default) and `"dft"` (Dynamic Fine-Tuning, as described in [this paper](https://huggingface.co/papers/2508.05629)). activation_offloading (`bool`, *optional*, defaults to `False`): Whether to offload the activations to the CPU. model_init_kwargsgh㈵>helpz$The initial learning rate for AdamW.)defaultmetadata learning_rate zLog every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, will be interpreted as ratio of total training steps. logging_stepsTzZIf True, use gradient checkpointing to save memory at the expense of slower backward pass.gradient_checkpointingNzWhether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if `fp16` is not set.bf16aCKeyword arguments for `AutoModelForCausalLM.from_pretrained`, used when the `model` argument of the `SFTTrainer` is provided as a string. If you're training a MoE architecture and want to include the load balancing/auxilliary loss as a part of the final loss, remember to set `output_router_logits=True` in this dictionary.acIf specified, sets the model's chat template. This can either be the path to a tokenizer (local directory or Hugging Face Hub model) or a direct path to a Jinja template file. When using a Jinja file, you must ensure that any special tokens referenced in the template are added to the tokenizer and that the model's embedding layer is resized accordingly.chat_template_pathtextz:Name of the column that contains text data in the dataset.dataset_text_fieldaTDictionary of optional keyword arguments for the dataset preparation. The only supported key is `skip_prepare_dataset`. If the model is a VLM, `skip_prepare_dataset` value is ignored. When the model is a VLM, `skip_prepare_dataset` is automatically treated as `True` regardless of the provided value, since preprocessing is done on the fly.dataset_kwargsz6Number of processes to use for processing the dataset.dataset_num_proczmToken used to indicate the end of a turn or sequence. If `None`, it defaults to `processing_class.eos_token`. eos_tokenzToken used for padding. If `None`, it defaults to `processing_class.pad_token`, or if that is also `None`, it falls back to `processing_class.eos_token`. pad_tokenizMaximum length of the tokenized sequence. Sequences longer than `max_length` are truncated fromthe right. If `None`, no truncation is applied. When packing is enabled, this value sets the sequence length. max_lengthFzWhether to group multiple sequences into fixed-length blocks to improve computational efficiency and reduce padding. Uses `max_length` to define sequence length.packingbfdzeStrategy for packing sequences. Can be either `'bfd'` (best-fit decreasing, default), or `'wrapped'`.packing_strategyaWhether to perform forward passes without padding by flattening all sequences in the batch into a single continuous sequence. This reduces memory usage by eliminating padding overhead. Currently, this is only supported with the FlashAttention 2 or 3, which can efficiently handle the flattened batch structure. When packing is enabled with strategy `'bfd'`, padding-free is enabled, regardless of the value of this parameter. padding_freezAIf set, the sequences will be padded to a multiple of this value.pad_to_multiple_ofzNWhether to pack the eval dataset. If `None`, uses the same value as `packing`. eval_packingaWhether to compute loss only on the completion part of the sequence. If set to `True`, loss is computed only on the completion, which is supported only for prompt-completion datasets. If `False`, loss is computed on the entire sequence. If `None` (default), the behavior depends on the dataset: loss is computed on the completion for prompt-completion datasets, and on the full sequence for language modeling datasets.completion_only_losszWhether to compute loss only on the assistant part of the sequence. If set to `True`, loss is computed only on the assistant responses, which is supported only for conversational datasets. If `False`, loss is computed on the entire sequence.assistant_only_lossnllzType of loss to use. Possible values are `"nll"` (negative log-likelihood, default) and `"dft"` (Dynamic Fine-Tuning, as described in https://huggingface.co/papers/2508.05629). loss_typez.Whether to offload the activations to the CPU.activation_offloadingc>URcUR(+O URUl[TU] 5 g)N)rfp16super __post_init__)self __class__s P/home/james-whalen/.local/lib/python3.13/site-packages/trl/trainer/sft_config.pyr*SFTConfig.__post_init__s*'+yy'8Odii  )r)(__name__ __module__ __qualname____firstlineno____doc__r_VALID_DICT_FIELDSrrfloat__annotations__rrboolrrr dictstrrrrrrintrrrrrrr r!r"r#r%r&r*__static_attributes__ __classcell__)r,s@r-r r s>M^+==AT@UU!@AM5! D M5$) p $D ! ! D(4.38 " 3xS#X/). G ) $VW05 6 0NHT#s(^,',RS'hsm % D  Ix} % M  Ix}!&  !J  O GT"  c '  L$ ).]^) $)jk$L(4. ,1 .  ,(4. !& ;  !  c Is#(JK#4   r/r N) dataclassesrrtypingrr transformersrr r/r-rBs/) * p !p  p r/