TPO Trainer ↗

Overview#

Triple Preference Optimization (TPO) was introduced in the paper Triple Preference Optimization: Achieving Better Alignment using a Single Step Optimization by Amir Saeidi, Shivanshu Verma, Aswin RRV, and Chitta Baral. TPO enhances the instruction-following and reasoning capabilities of large language models in a single training step, starting from a pre-trained or instruction-tuned model.

The abstract from the paper is the following:

Reinforcement Learning with Human Feedback (RLHF) enhances the alignment of Large Language Models (LLMs). However, its limitations have led to the development of Direct Preference Optimization (DPO), an RL-free approach designed to overcome these shortcomings. While studies have shown that DPO improves instruction-following capabilities, it negatively impacts the reasoning ability of LLMs. Additionally, DPO is highly sensitive to judgment noise in preference datasets and the size of the training set. Although several modifications to DPO have been proposed, they still fail to fully resolve these issues. To address these limitations, we propose Triple Preference Optimization (TPO), a new preference learning method designed to enhance both reasoning and instruction-following abilities through one-step optimization. We compare TPO against DPO and its recent variants using state-of-the-art training setups, including both base and instructiontuned models such as Mistral and Llama 3. Our evaluation covers a comprehensive range of chat-based and reasoning benchmarks. The results demonstrate that TPO achieves significant improvements over existing methods without substantially increasing response length across different dataset sizes. Specifically, TPO outperforms DPO and SimPO by up to 7.0% and 7.3% points on Arena-Hard, 12.2% and 13.3% points on MixEval-Hard, 10.4% and 10.1% points on MMLU-Pro, and 19.0% and 19.2% points on GSM8K, respectively. Furthermore, TPO achieves these improvements while requiring less data than DPO.

This post-training method was contributed by Kashif Rasul.

Quick start#

This example demonstrates how to train a model using the TPO method. We use the Qwen 3 0.6B model as the base model. TPO requires a triple-preference dataset (prompt, chosen, rejected, reference) — see Expected dataset type below.

Below is the script to train the model:

# train_tpo.py
from datasets import load_dataset
from trl.experimental.tpo import TPOConfig, TPOTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
train_dataset = load_dataset("tpo-alignment/triple-preference-ultrafeedback-40K", split="train")

training_args = TPOConfig(output_dir="Qwen3-0.6B-TPO")
trainer = TPOTrainer(model=model, args=training_args, processing_class=tokenizer, train_dataset=train_dataset)
trainer.train()

Execute the script using the following command:

accelerate launch train_tpo.py

Expected dataset type and format#

TPO requires a triple-preference dataset: each example must contain a prompt, a chosen (preferred) completion, a rejected (dispreferred) completion and a reference (gold) completion. The experimental.tpo.TPOTrainer supports both conversational and standard dataset formats. When provided with a conversational dataset, the trainer will automatically apply the chat template to the dataset.

# Standard format
triple_preference_example = {
    "prompt": "The sky is",
    "reference": " a beautiful shade of blue.",  # gold response (used for the NLL term)
    "chosen": " blue.",
    "rejected": " green.",
}

# Conversational format
triple_preference_example = {
    "prompt": [{"role": "user", "content": "What color is the sky?"}],
    "reference": [{"role": "assistant", "content": "It is a beautiful shade of blue."}],
    "chosen": [{"role": "assistant", "content": "It is blue."}],
    "rejected": [{"role": "assistant", "content": "It is green."}],
}

The reference response is typically the highest-quality completion available for the prompt; in the original TPO paper it is taken from the response with the highest score in UltraFeedback, with the second-highest used as the chosen completion and the lowest as the rejected completion.

Example script#

We provide an example script to train a model using the TPO method. The script is available at trl/experimental/tpo/tpo.py.

To test the TPO script with the Qwen 3 0.6B model on a triple-preference dataset, run the following command:

accelerate launch trl/experimental/tpo/tpo.py \
    --model_name_or_path Qwen/Qwen3-0.6B \
    --dataset_name tpo-alignment/triple-preference-ultrafeedback-40K \
    --beta 0.01 \
    --tpo_alpha 1.0 \
    --learning_rate 5e-7 \
    --num_train_epochs 1 \
    --output_dir Qwen3-0.6B-TPO

Looking deeper into the TPO method#

Triple Preference Optimization (TPO) extends preference-based alignment from pairs to triples (y_gold, y_chosen, y_rejected). The model is jointly optimized with two objectives in a single step:

1. A contrastive loss between the chosen and rejected completions, similar in spirit to DPO/SimPO but computed directly from the policy log-probabilities (no separate reference policy is required).
2. A supervised negative log-likelihood (NLL) loss on the gold (reference) completion, weighted by tpo_alpha. This term replaces the standalone SFT stage typically required before DPO.

The total TPO loss is:

$$ \mathcal{L}{\mathrm{TPO}}(\theta) = \mathcal{L}{\mathrm{contrast}}(\theta) + \alpha \cdot \mathcal{L}{\mathrm{NLL}}(\theta; y{\text{gold}}) $$

where \( \alpha \) is tpo_alpha and \( \mathcal{L}_{\mathrm{contrast}} \) is selected via loss_type.

Loss types#

Setting tpo_alpha=0.0 disables the NLL term entirely (the reference response is then unused, and the corresponding cross-entropy is skipped to save compute).

Logged metrics#

While training and evaluating we record the following metrics:

• loss: The total TPO loss (contrastive + tpo_alpha × NLL) averaged over the current logging interval.
• entropy: The average entropy of the model’s predicted token distribution over completion tokens.
• mean_token_accuracy: The proportion of completion tokens for which the model’s top-1 prediction matches the chosen completion.
• num_tokens: The total number of tokens processed so far.
• logits/chosen: The average logit values assigned by the model to the tokens in the chosen completion.
• logits/rejected: The average logit values assigned by the model to the tokens in the rejected completion.
• logps/chosen: The average log-probability assigned by the model to the chosen completion.
• logps/rejected: The average log-probability assigned by the model to the rejected completion.
• rewards/chosen: The average implicit reward computed for the chosen completion, defined as \( \beta \log \pi_{\theta}(y^{+}!\mid x) \).
• rewards/rejected: The average implicit reward computed for the rejected completion, defined as \( \beta \log \pi_{\theta}(y^{-}!\mid x) \).
• rewards/margins: The average implicit reward margin between the chosen and rejected completions.
• rewards/accuracies: The proportion of examples where the implicit reward for the chosen completion is higher than that for the rejected completion.

TPOTrainer[[trl.experimental.tpo.TPOTrainer]]#

trl.experimental.tpo.TPOTrainer[[trl.experimental.tpo.TPOTrainer]]#

Source

Trainer for Triple Preference Optimization (TPO) method. This algorithm was initially proposed in the paper Triple Preference Optimization: Achieving Better Alignment using a Single Step Optimization. This class is a wrapper around the Trainer class and inherits all of its attributes and methods.

traintrl.experimental.tpo.TPOTrainer.trainhttps://github.com/huggingface/trl/blob/v1.5.1/transformers/trainer.py#L1325[{“name”: “resume_from_checkpoint”, “val”: “: str | bool | None = None”}, {“name”: “trial”, “val”: “: optuna.Trial | dict[str, Any] | None = None”}, {“name”: “ignore_keys_for_eval”, “val”: “: list[str] | None = None”}]- resume_from_checkpoint (str or bool, optional) – If a str, local path to a saved checkpoint as saved by a previous instance of Trainer. If a bool and equals True, load the last checkpoint in args.output_dir as saved by a previous instance of Trainer. If present, training will resume from the model/optimizer/scheduler states loaded here.

• trial (optuna.Trial or dict[str, Any], optional) – The trial run or the hyperparameter dictionary for hyperparameter search.
• ignore_keys_for_eval (list[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions for evaluation during the training.0~trainer_utils.TrainOutputObject containing the global step count, training loss, and metrics.

Main training entry point.

Parameters:

model (str or PreTrainedModel or PeftModel) : Model to be trained. Can be either: - A string, being the model id of a pretrained model hosted inside a model repo on huggingface.co, or a path to a directory containing model weights saved using save_pretrained, e.g., './my_model_directory/'. The model is loaded using <ModelArchitecture>.from_pretrained (where <ModelArchitecture> is derived from the model config) with the keyword arguments in args.model_init_kwargs. - A PreTrainedModel object. Only causal language models are supported. - A PeftModel object. Only causal language models are supported.

args (experimental.tpo.TPOConfig, optional) : Configuration for this trainer. If None, a default configuration is used.

data_collator (DataCollator, optional) : Function to use to form a batch from a list of elements of the processed train_dataset or eval_dataset. Will default to DataCollatorForTriplePreference. Custom collators must truncate sequences before padding; the trainer does not apply post-collation truncation.

train_dataset (Dataset or IterableDataset) : Dataset to use for training. TPO requires a triple-preference dataset: each sample must contain a "chosen", a "rejected" and a "reference" (gold) completion. The format of the samples can be either: - Standard: Each sample contains plain text. - Conversational: Each sample contains structured messages (e.g., role and content).

eval_dataset (Dataset, IterableDataset or dict[str, Dataset | IterableDataset]) : Dataset to use for evaluation. It must meet the same requirements as train_dataset.

processing_class (PreTrainedTokenizerBase, optional) : Processing class used to process the data. If None, the processing class is loaded from the model’s name with from_pretrained. A padding token, tokenizer.pad_token, must be set. If the processing class has not set a padding token, tokenizer.eos_token will be used as the default.

compute_metrics (Callable[[EvalPrediction], dict], optional) : The function that will be used to compute metrics at evaluation. Must take a EvalPrediction and return a dictionary string to metric values.

callbacks (list of TrainerCallback, optional) : List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed in here. If you want to remove one of the default callbacks used, use the remove_callback method.

optimizers (tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None], optional, defaults to (None, None)) : A tuple containing the optimizer and the scheduler to use. Will default to an instance of AdamW on your model and a scheduler given by get_linear_schedule_with_warmup controlled by args.

peft_config (PeftConfig, optional) : PEFT configuration used to wrap the model. If None, the model is not wrapped.

Returns:

~trainer_utils.TrainOutput

Object containing the global step count, training loss, and metrics.

save_model[[trl.experimental.tpo.TPOTrainer.save_model]]#

Source

Will save the model, so you can reload it using from_pretrained().

Will only save from the main process.

push_to_hub[[trl.experimental.tpo.TPOTrainer.push_to_hub]]#

Source

Upload self.model and self.processing_class to the 🤗 model hub on the repo self.args.hub_model_id.

Parameters:

commit_message (str, optional, defaults to "End of training") : Message to commit while pushing.

blocking (bool, optional, defaults to True) : Whether the function should return only when the git push has finished.

token (str, optional, defaults to None) : Token with write permission to overwrite Trainer’s original args.

revision (str, optional) : The git revision to commit from. Defaults to the head of the “main” branch.

kwargs (dict[str, Any], optional) : Additional keyword arguments passed along to ~Trainer.create_model_card.

Returns:

The URL of the repository where the model was pushed if blocking=False, or a Future object tracking the progress of the commit if blocking=True.

TPOConfig[[trl.experimental.tpo.TPOConfig]]#

trl.experimental.tpo.TPOConfig[[trl.experimental.tpo.TPOConfig]]#

Source

Configuration class for the experimental.tpo.TPOTrainer.

This class includes only the parameters that are specific to TPO training. For a full list of training arguments, please refer to the TrainingArguments documentation. Note that default values in this class may differ from those in TrainingArguments.

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

These parameters have default values different from TrainingArguments:

• logging_steps: Defaults to 10 instead of 500.
• gradient_checkpointing: Defaults to True instead of False.
• bf16: Defaults to True if fp16 is not set, instead of False.
• learning_rate: Defaults to 5e-7 instead of 5e-5.