KTO Trainer

TRL supports the Kahneman-Tversky Optimization (KTO) Trainer for aligning language models with binary feedback data (e.g., upvote/downvote), as described in the paper by Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. For a full example have a look at examples/scripts/kto.py.

Depending on how good your base model is, you may or may not need to do SFT before KTO. This is different from standard RLHF and DPO, which always require SFT.

Expected dataset format

The KTO trainer expects a very specific format for the dataset as it does not require pairwise preferences. Since the model will be trained to directly optimize examples that consist of a prompt, model completion, and a label to indicate whether the completion is “good” or “bad”, we expect a dataset with the following columns:

prompt
completion
label

for example:

kto_dataset_dict = {
    "prompt": [
        "Hey, hello",
        "How are you",
        "What is your name?",
        "What is your name?",
        "Which is the best programming language?",
        "Which is the best programming language?",
        "Which is the best programming language?",
    ],
    "completion": [
        "hi nice to meet you",
        "leave me alone",
        "I don't have a name",
        "My name is Mary",
        "Python",
        "C++",
        "Java",
    ],
    "label": [
        True,
        False,
        False,
        True,
        True,
        False,
        False,
    ],
}

where the prompt contains the context inputs, completion contains the corresponding responses and label contains the corresponding flag that indicates if the generated completion is desired (True) or undesired (False). A prompt can have multiple responses and this is reflected in the entries being repeated in the dictionary’s value arrays. It is required that the dataset contains at least one desirable and one undesirable completion.

Expected model format

The KTO trainer expects a model of AutoModelForCausalLM, compared to PPO that expects AutoModelForCausalLMWithValueHead for the value function.

Using the KTOTrainer

For a detailed example have a look at the examples/scripts/kto.py script. At a high level we need to initialize the KTOTrainer with a model we wish to train and a reference ref_model which we will use to calculate the implicit rewards of the preferred and rejected response.

The beta refers to the hyperparameter of the implicit reward, and the dataset contains the 3 entries listed above. Note that the model and ref_model need to have the same architecture (ie decoder only or encoder-decoder).

The desirable_weight and undesirable_weight refer to the weights placed on the losses for desirable/positive and undesirable/negative examples. By default, they are both 1. However, if you have more of one or the other, then you should upweight the less common type such that the ratio of (desirable_weight number of positives) to (undesirable_weight number of negatives) is in the range 1:1 to 4:3.

training_args = KTOConfig(
    beta=0.1,
    desirable_weight=1.0,
    undesirable_weight=1.0,
)

kto_trainer = KTOTrainer(
    model,
    model_ref,
    args=training_args,
    train_dataset=train_dataset,
    tokenizer=tokenizer,
)

After this one can then call:

kto_trainer.train()

Loss Functions

Given the binary signal data indicating whether a completion is desirable or undesirable for a prompt, we can optimize an implicit reward function that aligns with the key principles of Kahneman-Tversky’s prospect theory, such as reference dependence, loss aversion, and diminishing sensitivity.

The BCO authors train a binary classifier whose logit serves as a reward so that the classifier maps {prompt, chosen completion} pairs to 1 and {prompt, rejected completion} pairs to 0. The KTOTrainer can be switched to this loss via the loss_type="bco" argument.

For Mixture of Experts Models: Enabling the auxiliary loss

MOEs are the most efficient if the load is about equally distributed between experts.
To ensure that we train MOEs similarly during preference-tuning, it is beneficial to add the auxiliary loss from the load balancer to the final loss.

This option is enabled by setting output_router_logits=True in the model config (e.g. MixtralConfig).
To scale how much the auxiliary loss contributes to the total loss, use the hyperparameter router_aux_loss_coef=... (default: 0.001).

KTOTrainer

class trl.KTOTrainer

< source >

( model: Union = None ref_model: Union = None args: KTOConfig = None train_dataset: Optional = None eval_dataset: Union = None tokenizer: Optional = None data_collator: Optional = None model_init: Optional = None callbacks: Optional = None optimizers: Tuple = (None, None) preprocess_logits_for_metrics: Optional = None peft_config: Optional = None compute_metrics: Optional = None model_adapter_name: Optional = None ref_adapter_name: Optional = None embedding_func: Optional = None embedding_tokenizer: Optional = None )

Parameters

model (transformers.PreTrainedModel) — The model to train, preferably an AutoModelForSequenceClassification.
ref_model (PreTrainedModelWrapper) — Hugging Face transformer model with a casual language modelling head. Used for implicit reward computation and loss. If no reference model is provided, the trainer will create a reference model with the same architecture as the model to be optimized.
args (KTOConfig) — The arguments to use for training.
train_dataset (datasets.Dataset) — The dataset to use for training.
eval_dataset (datasets.Dataset) — The dataset to use for evaluation.
tokenizer (transformers.PreTrainedTokenizerBase) — The tokenizer to use for training. This argument is required if you want to use the default data collator.
data_collator (transformers.DataCollator, optional, defaults to None) — The data collator to use for training. If None is specified, the default data collator (DPODataCollatorWithPadding) will be used which will pad the sequences to the maximum length of the sequences in the batch, given a dataset of paired sequences.
model_init (Callable[[], transformers.PreTrainedModel]) — The model initializer to use for training. If None is specified, the default model initializer will be used.
callbacks (List[transformers.TrainerCallback]) — The callbacks to use for training.
optimizers (Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]) — The optimizer and scheduler to use for training.
preprocess_logits_for_metrics (Callable[[torch.Tensor, torch.Tensor], torch.Tensor]) — The function to use to preprocess the logits before computing the metrics.
peft_config (Dict, defaults to None) — The PEFT configuration to use for training. If you pass a PEFT configuration, the model will be wrapped in a PEFT model.
disable_dropout (bool, defaults to True) — Whether or not to disable dropouts in model and ref_model.
compute_metrics (Callable[[EvalPrediction], Dict], optional) — The function to use to compute the metrics. Must take a EvalPrediction and return a dictionary string to metric values.
model_adapter_name (str, defaults to None) — Name of the train target PEFT adapter, when using LoRA with multiple adapters.
ref_adapter_name (str, defaults to None) — Name of the reference PEFT adapter, when using LoRA with multiple adapters.

Initialize KTOTrainer.

bco_loss

< source >

( policy_chosen_logps: FloatTensor policy_rejected_logps: FloatTensor reference_chosen_logps: FloatTensor reference_rejected_logps: FloatTensor chosen_embeddings: Optional rejected_embeddings: Optional ) → A tuple of four tensors

Returns

A tuple of four tensors

(losses, chosen_rewards, rejected_rewards, KL). The losses tensor contains the KTO loss for each example in the batch. The chosen_rewards and rejected_rewards tensors contain the rewards for the chosen and rejected responses, respectively. The delta value contains the moving average of all implicit rewards.

Compute the BCO loss for a batch of policy and reference model log probabilities.

compute_reference_log_probs

< source >

( padded_batch: Dict )

Computes log probabilities of the reference model for a single padded batch of a KTO specific dataset.

evaluation_loop

< source >

( dataloader: DataLoader description: str prediction_loss_only: Optional = None ignore_keys: Optional = None metric_key_prefix: str = 'eval' )

Overriding built-in evaluation loop to store metrics for each batch. Prediction/evaluation loop, shared by Trainer.evaluate() and Trainer.predict().

Works both with or without labels.

get_batch_logps

< source >

( logits: FloatTensor labels: LongTensor average_log_prob: bool = False label_pad_token_id: int = -100 is_encoder_decoder: bool = False )

Compute the log probabilities of the given labels under the given logits.

get_batch_loss_metrics

< source >

( model batch: Dict )

Compute the KTO loss and other metrics for the given batch of inputs for train or test.

get_batch_samples

< source >

( model batch: Dict )

Generate samples from the model and reference model for the given batch of inputs.

get_eval_dataloader

< source >

( eval_dataset: Optional = None )

Parameters

eval_dataset (torch.utils.data.Dataset, optional) — If provided, will override self.eval_dataset. If it is a Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement __len__.

Returns the evaluation ~torch.utils.data.DataLoader.

Subclass of transformers.src.transformers.trainer.get_eval_dataloader to precompute ref_log_probs.

get_train_dataloader

< source >

( )

Returns the training ~torch.utils.data.DataLoader.

Subclass of transformers.src.transformers.trainer.get_train_dataloader to precompute ref_log_probs.

kto_loss

< source >

( policy_chosen_logps: FloatTensor policy_rejected_logps: FloatTensor policy_KL_logps: FloatTensor reference_chosen_logps: FloatTensor reference_rejected_logps: FloatTensor reference_KL_logps: FloatTensor ) → A tuple of four tensors

Returns

A tuple of four tensors

(losses, chosen_rewards, rejected_rewards, KL). The losses tensor contains the KTO loss for each example in the batch. The chosen_rewards and rejected_rewards tensors contain the rewards for the chosen and rejected responses, respectively. The KL tensor contains the detached KL divergence estimate between the policy and reference models.

Compute the KTO loss for a batch of policy and reference model log probabilities.

log

< source >

( logs: Dict )

Parameters

logs (Dict[str, float]) — The values to log.

Log logs on the various objects watching training, including stored metrics.

null_ref_context

< source >

( )

Context manager for handling null reference model (that is, peft adapter manipulation).

KTOConfig

class trl.KTOConfig

< source >

( output_dir: str overwrite_output_dir: bool = False do_train: bool = False do_eval: bool = False do_predict: bool = False eval_strategy: Union = 'no' prediction_loss_only: bool = False per_device_train_batch_size: int = 8 per_device_eval_batch_size: int = 8 per_gpu_train_batch_size: Optional = None per_gpu_eval_batch_size: Optional = None gradient_accumulation_steps: int = 1 eval_accumulation_steps: Optional = None eval_delay: Optional = 0 torch_empty_cache_steps: Optional = None learning_rate: float = 5e-05 weight_decay: float = 0.0 adam_beta1: float = 0.9 adam_beta2: float = 0.999 adam_epsilon: float = 1e-08 max_grad_norm: float = 1.0 num_train_epochs: float = 3.0 max_steps: int = -1 lr_scheduler_type: Union = 'linear' lr_scheduler_kwargs: Union = <factory> warmup_ratio: float = 0.0 warmup_steps: int = 0 log_level: Optional = 'passive' log_level_replica: Optional = 'warning' log_on_each_node: bool = True logging_dir: Optional = None logging_strategy: Union = 'steps' logging_first_step: bool = False logging_steps: float = 500 logging_nan_inf_filter: bool = True save_strategy: Union = 'steps' save_steps: float = 500 save_total_limit: Optional = None save_safetensors: Optional = True save_on_each_node: bool = False save_only_model: bool = False restore_callback_states_from_checkpoint: bool = False no_cuda: bool = False use_cpu: bool = False use_mps_device: bool = False seed: int = 42 data_seed: Optional = None jit_mode_eval: bool = False use_ipex: bool = False bf16: bool = False fp16: bool = False fp16_opt_level: str = 'O1' half_precision_backend: str = 'auto' bf16_full_eval: bool = False fp16_full_eval: bool = False tf32: Optional = None local_rank: int = -1 ddp_backend: Optional = None tpu_num_cores: Optional = None tpu_metrics_debug: bool = False debug: Union = '' dataloader_drop_last: bool = False eval_steps: Optional = None dataloader_num_workers: int = 0 dataloader_prefetch_factor: Optional = None past_index: int = -1 run_name: Optional = None disable_tqdm: Optional = None remove_unused_columns: Optional = True label_names: Optional = None load_best_model_at_end: Optional = False metric_for_best_model: Optional = None greater_is_better: Optional = None ignore_data_skip: bool = False fsdp: Union = '' fsdp_min_num_params: int = 0 fsdp_config: Union = None fsdp_transformer_layer_cls_to_wrap: Optional = None accelerator_config: Union = None deepspeed: Union = None label_smoothing_factor: float = 0.0 optim: Union = 'adamw_torch' optim_args: Optional = None adafactor: bool = False group_by_length: bool = False length_column_name: Optional = 'length' report_to: Union = None ddp_find_unused_parameters: Optional = None ddp_bucket_cap_mb: Optional = None ddp_broadcast_buffers: Optional = None dataloader_pin_memory: bool = True dataloader_persistent_workers: bool = False skip_memory_metrics: bool = True use_legacy_prediction_loop: bool = False push_to_hub: bool = False resume_from_checkpoint: Optional = None hub_model_id: Optional = None hub_strategy: Union = 'every_save' hub_token: Optional = None hub_private_repo: bool = False hub_always_push: bool = False gradient_checkpointing: bool = False gradient_checkpointing_kwargs: Union = None include_inputs_for_metrics: bool = False eval_do_concat_batches: bool = True fp16_backend: str = 'auto' evaluation_strategy: Union = None push_to_hub_model_id: Optional = None push_to_hub_organization: Optional = None push_to_hub_token: Optional = None mp_parameters: str = '' auto_find_batch_size: bool = False full_determinism: bool = False torchdynamo: Optional = None ray_scope: Optional = 'last' ddp_timeout: Optional = 1800 torch_compile: bool = False torch_compile_backend: Optional = None torch_compile_mode: Optional = None dispatch_batches: Optional = None split_batches: Optional = None include_tokens_per_second: Optional = False include_num_input_tokens_seen: Optional = False neftune_noise_alpha: Optional = None optim_target_modules: Union = None batch_eval_metrics: bool = False eval_on_start: bool = False eval_use_gather_object: Optional = False max_length: Optional = None max_prompt_length: Optional = None max_completion_length: Optional = None beta: float = 0.1 desirable_weight: Optional = 1.0 undesirable_weight: Optional = 1.0 label_pad_token_id: int = -100 padding_value: int = None truncation_mode: str = 'keep_end' generate_during_eval: bool = False is_encoder_decoder: Optional = None precompute_ref_log_probs: bool = False model_init_kwargs: Optional = None ref_model_init_kwargs: Optional = None dataset_num_proc: Optional = None loss_type: Literal = 'kto' prompt_sample_size: int = 1024 min_density_ratio: float = 0.5 max_density_ratio: float = 10.0 )

Parameters

max_length (int, optional, defaults to None) — The maximum length of the sequences in the batch. This argument is required if you want to use the default data collator.
max_prompt_length (int, optional, defaults to None) — The maximum length of the prompt. This argument is required if you want to use the default data collator.
max_completion_length (int, optional, defaults to None) — The maximum length of the target. This argument is required if you want to use the default data collator and your model is an encoder-decoder.
beta (float, defaults to 0.1) — The beta factor in KTO loss. Higher beta means less divergence from the initial policy.
desirable_weight (float, optional, defaults to 1.0) — The desirable losses are weighed by this factor to counter unequal number of desirable and undesirable paris.
undesirable_weight (float, optional, defaults to 1.0) — The undesirable losses are weighed by this factor to counter unequal number of desirable and undesirable pairs.
label_pad_token_id (int, defaults to -100) — The label pad token id. This argument is required if you want to use the default data collator.
padding_value (int, defaults to 0) — The padding value if it is different to the tokenizer’s pad_token_id.
truncation_mode (str, defaults to keep_end) — The truncation mode to use, either keep_end or keep_start. This argument is required if you want to use the default data collator.
generate_during_eval (bool, defaults to False) — Whether to sample and log generations during evaluation step.
is_encoder_decoder (Optional[bool], optional, defaults to None) — If no model is provided, we need to know if the model_init returns an encoder-decoder.
precompute_ref_log_probs (bool, defaults to False) — Flag to precompute reference model log probabilities for training and evaluation datasets. This is useful if you want to train without the reference model and reduce the total GPU memory needed. model_init_kwargs — (Optional[Dict], optional): Dict of Optional kwargs to pass when instantiating the model from a string. ref_model_init_kwargs — (Optional[Dict], optional): Dict of Optional kwargs to pass when instantiating the ref model from a string. dataset_num_proc — (Optional[int], optional, defaults to None): Number of processes to use for processing the datasets. loss_type — (Literal["kto", "bco"], optional): The type of loss to use. Either "kto" the default KTO loss, "bco" loss from BCO paper. prompt_sample_size — (int, defaults to 1024): Number of prompts that are fed to density ratio classifier. min_density_ratio — (float, defaults to 0.5): The minimum value of the density ratio. The estimated density ratio is clamped to this value. max_density_ratio — (float, defaults to 10.0): The maximum value of the density ratio. The estimated density ratio is clamped to this value.

KTOConfig collects all training arguments related to the KTOTrainer class.

Using HfArgumentParser we can turn this class into argparse arguments that can be specified on the command line.

< > Update on GitHub