Online DPO Trainer

Overview

Online DPO was proposed in Direct Language Model Alignment from Online AI Feedback by Shangmin Guo, Biao Zhang, Tianlin Liu, Tianqi Liu, Misha Khalman, Felipe Llinares, Alexandre Rame, Thomas Mesnard, Yao Zhao, Bilal Piot, Johan Ferret, and Mathieu Blondel.

The abstract from the paper is the following:

Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged as efficient alternatives to reinforcement learning from human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus the feedback is purely offline. Moreover, responses in these datasets are often sampled from a language model distinct from the one being aligned, and since the model evolves over training, the alignment phase is inevitably off-policy. In this study, we posit that online feedback is key and improves DAP methods. Our method, online AI feedback (OAIF), uses an LLM as annotator: on each training iteration, we sample two responses from the current model and prompt the LLM annotator to choose which one is preferred, thus providing online feedback. Despite its simplicity, we demonstrate via human evaluation in several tasks that OAIF outperforms both offline DAP and RLHF methods. We further show that the feedback leveraged in OAIF is easily controllable, via instruction prompts to the LLM annotator.

This post-training method was contributed by Michael Noukhovitch, Shengyi Costa Huang, Quentin Gallouédec, and Edward Beeching.

Quick start

This example demonstrates how to train a model using the online DPO method. We use the Qwen 0.5B model as the base model and PairRMJudge as a judge. We use the prompts from the UltraFeedback dataset. You can view the prompts in the dataset here:

Below is the script to train the model:

# train_online_dpo.py
from datasets import load_dataset
from trl import OnlineDPOConfig, OnlineDPOTrainer, PairRMJudge
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
judge = PairRMJudge()
train_dataset = load_dataset("trl-lib/ultrafeedback-prompt", split="train")

training_args = OnlineDPOConfig(output_dir="Qwen2-0.5B-OnlineDPO", logging_steps=10)
trainer = OnlineDPOTrainer(
    model=model, judge=judge, args=training_args, processing_class=tokenizer, train_dataset=train_dataset
)
trainer.train()

Execute the script using the following command:

accelerate launch train_online_dpo.py

Distributed across 8 GPUs, the training takes approximately 1 hour. You can verify the training progress by checking the reward graph. An increasing trend in both the reward for rejected and chosen completions indicates that the model is improving and generating better responses over time.

To see how the trained model performs, you can use the TRL Chat CLI.

$ trl chat --model_name_or_path trl-lib/Qwen2-0.5B-OnlineDPO
<quentin_gallouedec>:
What is the best programming language?

<trl-lib/Qwen2-0.5B-OnlineDPO>:
The best programming language depends on your specific needs and priorities. Some people prefer imperative programming languages (like Haskell or Lisp), while others prefer functional programming languages (like Scala or Python). It's important to consider your work style, programming environment, and project requirements when choosing a programming language.

Expected dataset type

Online DPO only requires a prompt-only dataset (unlike offline DPO, that expects preference dataset). The OnlineDPOTrainer supports both conversational and standard dataset format. When provided with a conversational dataset, the trainer will automatically apply the chat template to the dataset.

Usage tips

Use a reward model

Instead of a judge, you can chose to use a reward model — see Reward Bench for a leaderboard of public models you can use. Below is a code example showing how to replace a judge with the trl-lib/Qwen2-0.5B-Reward model:

- from trl import PairRMJudge
+ from transformers import AutoModelForSequenceClassification

- judge = PairRMJudge()
+ reward_model = AutoModelForSequenceClassification.from_pretrained("trl-lib/Qwen2-0.5B-Reward", num_labels=1)
+ reward_tokenizer = AutoTokenizer.from_pretrained("trl-lib/Qwen2-0.5B-Reward")

  trainer = OnlineDPOTrainer(
      ...
-     judge=judge,
+     reward_model=reward_model,
+     reward_processing_class=reward_tokenizer,
      ...
  )

Encourage EOS token generation

When using a reward model, we may want the model to generate completions within a given length. During training, the model will generate completions up to the maximum length specified in the max_new_tokens argument of OnlineDPOConfig. If you want to penalize the model for not generating an EOS token before reaching the maximum length, you can use the missing_eos_penalty argument of OnlineDPOConfig:

training_args = OnlineDPOConfig(..., max_new_tokens=128, missing_eos_penalty=1.0)

Logging Completions

To better understand your model’s behavior during training, you can log sample completions periodically using the LogCompletionsCallback.

trainer = OnlineDPOTrainer(..., eval_dataset=eval_dataset)
completions_callback = LogCompletionsCallback(trainer, num_prompts=8)
trainer.add_callback(completions_callback)

This callback logs the model’s generated completions directly to Weights & Biases.

Logged Completions

Example script

We provide an example script to train a model using the online DPO method. The script is available in examples/scripts/dpo_online.py

To test the online DPO script with the Qwen2.5 0.5B model on the UltraFeedback dataset, run the following command:

python examples/scripts/dpo_online.py \
    --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
    --judge pair_rm \
    --dataset_name trl-lib/ultrafeedback-prompt \
    --learning_rate 5.0e-7 \
    --logging_steps 25 \
    --output_dir Qwen2.5-0.5B-Online-DPO-PairRM \
    --warmup_ratio 0.1 \
    --push_to_hub

Logged metrics

The logged metrics are as follows. Here is an example tracked run at Weights and Biases

objective/kl: The mean Kullback-Leibler (KL) divergence between the current model and reference model.
objective/entropy: The mean entropy of the model, indicating the randomness of the actions chosen by the model.
objective/non_score_reward: The mean reward from non-score-related sources, basically beta * kl.sum(1), where beta is the KL penalty coefficient and kl is the per-token KL divergence.
objective/rlhf_reward: The mean RLHF reward, which is scores - non_score_reward. The rlhf_reward is the ultimate objective of online DPO training. If training works as intended, this metric should keep going up.
objective/scores: The mean scores returned by the reward model.
objective/scores_margin: The mean score margin (according to the external reward model) between the chosen and rejected completions.
rewards/chosen: The mean reward (according to online DPO’s implicit reward model)of the chosen completions.
rewards/rejected: The mean reward (according to online DPO’s implicit reward model) of the rejected completions.
rewards/accuracies: The accuracies of the online DPO’s implicit reward model.
rewards/margins: The mean reward margin (according to online DPO’s implicit reward model) between the chosen and rejected completions.
logps/chosen: The mean log probabilities of the chosen completions.
logps/rejected: The mean log probabilities of the rejected completions.
val/contain_eos_token: The fraction of completions which contain an EOS token.
beta: The parameter that controls the weight of the loss term representing the deviation from the reference model. Typically fixed, but can be made dynamic by passing a list to OnlineDPOConfig.

Benchmark experiments

To validate the online DPO implementation works, we ran experiments with the Pythia 1B, 2.8B, and 6.9B models on a single node of 8 x H100s. Here are the commands we used to run the experiments. We take the SFT / RM models directly from The N+ Implementation Details of RLHF with PPO: A Case Study on TL;DR Summarization.

# 1B Online DPO experiment
accelerate launch --config_file examples/accelerate_configs/multi_gpu.yaml \
    examples/scripts/dpo_online.py \
    --model_name_or_path trl-lib/pythia-1b-deduped-tldr-sft  \
    --reward_model_path trl-lib/pythia-1b-deduped-tldr-rm \
    --dataset_name trl-lib/tldr \
    --learning_rate 5.0e-7 \
    --output_dir pythia-1b-deduped-tldr-online-dpo \
    --beta 0.1 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 2 \
    --num_train_epochs 3 \
    --max_new_tokens 53 \
    --warmup_ratio 0.1 \
    --missing_eos_penalty 1.0 \
    --logging_steps 20 \
    --save_steps 0.1 \
    --push_to_hub

# 2.8B Online DPO experiment
accelerate launch --config_file examples/accelerate_configs/deepspeed_zero2.yaml \
    examples/scripts/dpo_online.py \
    --model_name_or_path trl-lib/pythia-2.8b-deduped-tldr-sft  \
    --reward_model_path trl-lib/pythia-2.8b-deduped-tldr-rm \
    --dataset_name trl-lib/tldr \
    --learning_rate 5.0e-7 \
    --output_dir pythia-2.8b-deduped-tldr-online-dpo \
    --beta 0.1 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 2 \
    --num_train_epochs 3 \
    --max_new_tokens 53 \
    --warmup_ratio 0.1 \
    --missing_eos_penalty 1.0 \
    --bf16 \
    --logging_steps 20 \
    --save_steps 0.1 \
    --push_to_hub

# 6.9B Online DPO experiment
accelerate launch --config_file examples/accelerate_configs/deepspeed_zero2.yaml \
    examples/scripts/dpo_online.py \
    --model_name_or_path trl-lib/pythia-6.9b-deduped-tldr-sft  \
    --reward_model_path trl-lib/pythia-6.9b-deduped-tldr-rm \
    --dataset_name trl-lib/tldr \
    --learning_rate 5.0e-7 \
    --output_dir pythia-6.9b-deduped-tldr-online-dpo \
    --beta 0.1 \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --num_train_epochs 3 \
    --max_new_tokens 53 \
    --warmup_ratio 0.1 \
    --missing_eos_penalty 1.0 \
    --bf16 \
    --gradient_checkpointing \
    --logging_steps 20 \
    --save_steps 0.1 \
    --push_to_hub

Checkpoints and experiment tracking are available at:

To evaluate, we use vLLM to load the checkpoints and GPT-4o mini as a judge model to evaluate the generated TL;DR against the reference TL;DR. For more information on how to use judges, see Judges.

$ python examples/scripts/evals/judge_tldr.py --model_name_or_path trl-lib/pythia-1b-deduped-tldr-sft --judge_model gpt-4o-mini --num_examples 1000
Model win rate: 33.00%
python examples/scripts/evals/judge_tldr.py --model_name_or_path trl-lib/pythia-6.9b-deduped-tldr-sft --judge_model gpt-4o-mini --num_examples 1000
Model win rate: 41.50%
python examples/scripts/evals/judge_tldr.py --model_name_or_path trl-lib/pythia-1b-deduped-tldr-online-dpo --judge_model gpt-4o-mini --num_examples 1000
Model win rate: 62.60%
python examples/scripts/evals/judge_tldr.py --model_name_or_path trl-lib/pythia-6.9b-deduped-tldr-online-dpo --judge_model gpt-4o-mini --num_examples 1000
Model win rate: 74.20%

We can then plot the RLHF scaling chart.

import matplotlib.pyplot as plt

results = {
    "SFT": {1.0e9: 0.21, 2.8e9: 0.27, 6.9e9: 0.316},
    "online-dpo": {1.0e9: 0.542, 2.8e9: 0.746, 6.9e9: 0.796},
    "offline-dpo": {1.0e9: 0.422, 2.8e9: 0.517, 6.9e9: 0.701},
}


plt.plot(results["SFT"].keys(), results["SFT"].values(), label="SFT", marker="o")
plt.plot(results["online-dpo"].keys(), results["online-dpo"].values(), label="Online-dpo with RM judge", marker="o")
plt.plot(results["offline-dpo"].keys(), results["offline-dpo"].values(), label="Offline-dpo", marker="o")
plt.axhline(y=0.5, color="black", linestyle="-.", label="Human reference summary")
plt.xscale("log")
plt.xlabel("Model size")
plt.ylabel("Win rate against reference summaries\n(according to GPT-4-0613)")
plt.title("DPO scaling by model size")
plt.legend()
plt.xlim(5e8, 1.2e10)
plt.xticks([1e9, 3e9, 1e10], ["1B", "3B", "10B"])
plt.grid(True, which="both", ls="--", c="0.7")
plt.tight_layout()
plt.show()

The online DPO checkpoint gets increasingly more win rate as we scale up the model sizes. This is a good sign that the online DPO implementation is working as intended.

OnlineDPOTrainer

class trl.OnlineDPOTrainer

< source >

( model: typing.Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module] ref_model: typing.Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module, NoneType] = None reward_model: typing.Union[transformers.modeling_utils.PreTrainedModel, torch.nn.modules.module.Module, NoneType] = None judge: typing.Optional[trl.trainer.judges.BasePairwiseJudge] = None args: typing.Optional[trl.trainer.online_dpo_config.OnlineDPOConfig] = None data_collator: typing.Optional[transformers.data.data_collator.DataCollator] = None train_dataset: typing.Union[datasets.arrow_dataset.Dataset, torch.utils.data.dataset.IterableDataset, ForwardRef('datasets.Dataset'), NoneType] = None eval_dataset: typing.Union[datasets.arrow_dataset.Dataset, dict[str, datasets.arrow_dataset.Dataset], ForwardRef('datasets.Dataset'), NoneType] = None processing_class: typing.Union[transformers.tokenization_utils_base.PreTrainedTokenizerBase, transformers.image_processing_utils.BaseImageProcessor, transformers.feature_extraction_utils.FeatureExtractionMixin, transformers.processing_utils.ProcessorMixin, NoneType] = None reward_processing_class: typing.Optional[transformers.tokenization_utils_base.PreTrainedTokenizerBase] = None peft_config: typing.Optional[dict] = None compute_metrics: typing.Optional[typing.Callable[[transformers.trainer_utils.EvalPrediction], dict]] = None callbacks: typing.Optional[list[transformers.trainer_callback.TrainerCallback]] = None optimizers: tuple = (None, None) preprocess_logits_for_metrics: typing.Optional[typing.Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None )

Parameters

model (transformers.PreTrainedModel or torch.nn.Module) — The model to train, preferably an AutoModelForCausalLM.
ref_model (transformers.PreTrainedModel or torch.nn.Module or None) — The reference model to use for training. If None is specified, the reference model will be created from the model.
reward_model (transformers.PreTrainedModel or torch.nn.Module or None) — The reward model to score completions with, preferably an AutoModelForSequenceClassification.
judge (BasePairwiseJudge) — The judge to use for pairwise comparison of model completions.
args (OnlineDPOConfig) — The online DPO config arguments to use for training.
data_collator (transformers.DataCollator) — 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.
train_dataset (datasets.Dataset) — The dataset to use for training.
eval_dataset (datasets.Dataset) — The dataset to use for evaluation.
processing_class (PreTrainedTokenizerBase or BaseImageProcessor or FeatureExtractionMixin or ProcessorMixin, optional) — Processing class used to process the data. If provided, will be used to automatically process the inputs for the model, and it will be saved along the model to make it easier to rerun an interrupted training or reuse the fine-tuned model.
peft_config (dict) — The peft config to use for training.
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.
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.

Initialize OnlineDPOTrainer.

create_model_card

< source >

( model_name: typing.Optional[str] = None dataset_name: typing.Optional[str] = None tags: typing.Union[str, list[str], NoneType] = None )

Parameters

model_name (str, optional, defaults to None) — The name of the model.
dataset_name (str, optional, defaults to None) — The name of the dataset used for training.
tags (str, list[str] or None, optional, defaults to None) — Tags to be associated with the model card.

Creates a draft of a model card using the information available to the Trainer.

tokenize_row

< source >

( feature is_encoder_decoder: bool tokenizer: PreTrainedTokenizerBase )

Tokenize a single row from a DPO specific dataset.

OnlineDPOConfig

class trl.OnlineDPOConfig

< source >

( output_dir: str overwrite_output_dir: bool = False do_train: bool = False do_eval: bool = False do_predict: bool = False eval_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = '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: typing.Optional[int] = None per_gpu_eval_batch_size: typing.Optional[int] = None gradient_accumulation_steps: int = 1 eval_accumulation_steps: typing.Optional[int] = None eval_delay: typing.Optional[float] = 0 torch_empty_cache_steps: typing.Optional[int] = None learning_rate: float = 5e-07 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: typing.Union[transformers.trainer_utils.SchedulerType, str] = 'linear' lr_scheduler_kwargs: typing.Union[dict, str, NoneType] = <factory> warmup_ratio: float = 0.0 warmup_steps: int = 0 log_level: typing.Optional[str] = 'passive' log_level_replica: typing.Optional[str] = 'warning' log_on_each_node: bool = True logging_dir: typing.Optional[str] = None logging_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps' logging_first_step: bool = False logging_steps: float = 500 logging_nan_inf_filter: bool = True save_strategy: typing.Union[transformers.trainer_utils.SaveStrategy, str] = 'steps' save_steps: float = 500 save_total_limit: typing.Optional[int] = None save_safetensors: typing.Optional[bool] = 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: typing.Optional[int] = 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: typing.Optional[bool] = None local_rank: int = -1 ddp_backend: typing.Optional[str] = None tpu_num_cores: typing.Optional[int] = None tpu_metrics_debug: bool = False debug: typing.Union[str, typing.List[transformers.debug_utils.DebugOption]] = '' dataloader_drop_last: bool = False eval_steps: typing.Optional[float] = None dataloader_num_workers: int = 0 dataloader_prefetch_factor: typing.Optional[int] = None past_index: int = -1 run_name: typing.Optional[str] = None disable_tqdm: typing.Optional[bool] = None remove_unused_columns: typing.Optional[bool] = True label_names: typing.Optional[typing.List[str]] = None load_best_model_at_end: typing.Optional[bool] = False metric_for_best_model: typing.Optional[str] = None greater_is_better: typing.Optional[bool] = None ignore_data_skip: bool = False fsdp: typing.Union[typing.List[transformers.trainer_utils.FSDPOption], str, NoneType] = '' fsdp_min_num_params: int = 0 fsdp_config: typing.Union[dict, str, NoneType] = None fsdp_transformer_layer_cls_to_wrap: typing.Optional[str] = None accelerator_config: typing.Union[dict, str, NoneType] = None deepspeed: typing.Union[dict, str, NoneType] = None label_smoothing_factor: float = 0.0 optim: typing.Union[transformers.training_args.OptimizerNames, str] = 'adamw_torch' optim_args: typing.Optional[str] = None adafactor: bool = False group_by_length: bool = False length_column_name: typing.Optional[str] = 'length' report_to: typing.Union[NoneType, str, typing.List[str]] = None ddp_find_unused_parameters: typing.Optional[bool] = None ddp_bucket_cap_mb: typing.Optional[int] = None ddp_broadcast_buffers: typing.Optional[bool] = 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: typing.Optional[str] = None hub_model_id: typing.Optional[str] = None hub_strategy: typing.Union[transformers.trainer_utils.HubStrategy, str] = 'every_save' hub_token: typing.Optional[str] = None hub_private_repo: typing.Optional[bool] = None hub_always_push: bool = False gradient_checkpointing: bool = False gradient_checkpointing_kwargs: typing.Union[dict, str, NoneType] = None include_inputs_for_metrics: bool = False include_for_metrics: typing.List[str] = <factory> eval_do_concat_batches: bool = True fp16_backend: str = 'auto' evaluation_strategy: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = None push_to_hub_model_id: typing.Optional[str] = None push_to_hub_organization: typing.Optional[str] = None push_to_hub_token: typing.Optional[str] = None mp_parameters: str = '' auto_find_batch_size: bool = False full_determinism: bool = False torchdynamo: typing.Optional[str] = None ray_scope: typing.Optional[str] = 'last' ddp_timeout: typing.Optional[int] = 1800 torch_compile: bool = False torch_compile_backend: typing.Optional[str] = None torch_compile_mode: typing.Optional[str] = None dispatch_batches: typing.Optional[bool] = None split_batches: typing.Optional[bool] = None include_tokens_per_second: typing.Optional[bool] = False include_num_input_tokens_seen: typing.Optional[bool] = False neftune_noise_alpha: typing.Optional[float] = None optim_target_modules: typing.Union[NoneType, str, typing.List[str]] = None batch_eval_metrics: bool = False eval_on_start: bool = False use_liger_kernel: typing.Optional[bool] = False eval_use_gather_object: typing.Optional[bool] = False average_tokens_across_devices: typing.Optional[bool] = False reward_model_path: typing.Optional[str] = None judge: typing.Optional[str] = None max_new_tokens: int = 64 temperature: float = 0.9 missing_eos_penalty: typing.Optional[float] = None beta: list = <factory> loss_type: typing.Literal['sigmoid', 'ipo'] = 'sigmoid' dataset_num_proc: typing.Optional[int] = None disable_dropout: bool = True )

Parameters

learning_rate (float, optional, defaults to 5e-7) — Initial learning rate for AdamW optimizer. The default value replaces that of TrainingArguments.
reward_model_path (Optional[str], optional, defaults to None) — Path to the reward model. Either judge or reward_model_path must be set, but not both.
judge (Optional[str], optional, defaults to None) — Name of the judge to use. Either judge or reward_model_path must be set, but not both.
max_new_tokens (int, optional, defaults to 64) — Maximum number of tokens to generate per completion.
temperature (float, optional, defaults to 0.9) — Temperature for sampling. The higher the temperature, the more random the completions.
missing_eos_penalty (Optional[float], optional, defaults to None) — Penalty applied to the score when the model fails to generate an EOS token. This is useful to encourage to generate completions shorter than the maximum length (max_new_tokens). The penalty must be a positive value.
beta (float or list[float], optional, defaults to 0.1) — Parameter controlling the deviation from the reference model. Higher β means less deviation from the reference model. For the IPO loss (loss_type="ipo"), β is the regularization parameter denoted by τ in the paper. If a list of floats is provided then the β is selected for each new epoch and the last β is used for the rest of the epochs.
loss_type (str, optional, defaults to "sigmoid") — Type of loss to use. Possible values are:
- "sigmoid": sigmoid loss from the original DPO paper.
- "ipo": IPO loss from the IPO paper.
dataset_num_proc (Optional[int], optional, defaults to None) — Number of processes to use for processing the dataset.
disable_dropout (bool, optional, defaults to True) — Whether to disable dropout in the model.

Configuration class for the OnlineDPOTrainer.

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

< > Update on GitHub