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.

The current implementation uses reward models for scoring completions — see Reward Bench for a leaderboard of public models you can use.

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

Usage tips

Make sure that the SFT model and reward model use the same chat template. Otherwise, you may find the model completions are scored incorrectly during training.

The basic API is as follows:

from datasets import Dataset
from trl import OnlineDPOConfig, OnlineDPOTrainer
from transformers import (
    AutoModelForCausalLM,
    AutoModelForSequenceClassification,
    AutoTokenizer,
)
NUM_DUMMY_SAMPLES = 100

tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM-135M-Instruct")
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
# The model to optimise
model = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM-135M-Instruct")
# The reference model to calculate the KL divergence against
ref_model = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM-135M-Instruct")
# The model to score completions with. In practice, you will need a reward model.
reward_model = AutoModelForSequenceClassification.from_pretrained("HuggingFaceTB/SmolLM-135M-Instruct", num_labels=1)

train_dataset = Dataset.from_dict(
    {"prompt": ["Q: Hi how are you? A:"] * NUM_DUMMY_SAMPLES})
eval_dataset = Dataset.from_dict(
    {"prompt": ["Q: What do you like to eat A:"] * NUM_DUMMY_SAMPLES})

args = OnlineDPOConfig(output_dir="online-dpo-model")
trainer = OnlineDPOTrainer(
    model=model,
    ref_model=ref_model,
    reward_model=reward_model,
    args=args,
    tokenizer=tokenizer,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)
trainer.train()

To test the online DPO script with 1B parameter models, run:

python 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-tldr-online-dpo \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 32 \
    --num_train_epochs 3 \
    --max_new_tokens 53 \
    --warmup_ratio 0.1 \
    --missing_eos_penalty 1.0 \
    --push_to_hub

Tips:

objective/rlhf_reward is the ultimate objective of online DPO training. If training works as intended, this metric should keep going up.
We recommend using the “EOS trick” via the --missing_eos_penalty argument, which subtracts from the rewards a fixed scalar penalty for completions that do not end with an EOS token. This can help the model learn to generate more coherent completions.

Expected dataset format

Unlike offline DPO, where one provides a dataset with chosen and rejected columns, online DPO only requires a dataset of prompts to generate the completions from. The OnlineDPOTrainer assumes that the dataset is preprocessed for model inference, so typically you will need to wrap your prompts in the messages format and then apply the chat template as follows:

def prepare_dataset(row):
    """Apply chat template to messages"""
    row["prompt"] = tokenizer.apply_chat_template(row["prompt"], tokenize=False, add_generation_prompt=True)
    return row

dataset = prepare_dataset(dataset)

Explanation of the 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 score - non_score_reward.
objective/scores: The mean scores returned by the reward model / environment.
objective/scores_margin: The mean score margin (according to the external reward model) between the chosen and rejected completions.
rewards/accuracies: The accuracies of the online DPO’s implicit reward model.
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/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.

What is my model doing exactly?

To help you understand what your model is doing, we periodically log some sample completions from the model via LogCompletionsCallback. You can find an example tracked run at Weights and Biases, which allows you to see the model’s response at different stages of training. By default we generate during training, but you can customize the number of prompts to generate for in LogCompletionsCallback.

Implementation details

Many online implementation details are borrowed from the PPOv2Trainer, which is itself based on the The N+ Implementation Details of RLHF with PPO: A Case Study on TL;DR Summarization.

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: Union = None ref_model: Union = None reward_model: Optional = None judge: Optional = None args: Optional = None data_collator: Optional = None train_dataset: Union = None eval_dataset: Union = None tokenizer: Optional = None model_init: Optional = None compute_metrics: Optional = None callbacks: Optional = None optimizers: Tuple = (None, None) preprocess_logits_for_metrics: Optional = None )

Parameters

model (transformers.PreTrainedModel) — The model to train, preferably an AutoModelForCausalLM.
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.
reward_model (transformers.PreTrainedModel) — 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.
tokenizer (transformers.PreTrainedTokenizerBase) — The tokenizer to use for training. This argument is required if you want to use the default data collator.
model_init (Callable[[], transformers.PreTrainedModel]) — The model initializer to use for training. If None is specified, the default model initializer will be used.
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.

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: 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 use_liger_kernel: Optional = False eval_use_gather_object: Optional = False reward_model_path: Optional = None max_new_tokens: int = 53 temperature: float = 0.9 missing_eos_penalty: Optional = None beta: float = 0.1 loss_type: Literal = 'sigmoid' dataset_num_proc: Optional = None )

Parameters

reward_model_path (Optional[str], optional, defaults to None) — Path to the reward model.
max_new_tokens (int, optional, defaults to 64) — The 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 when the model fails to generate an EOS token.
beta (float, optional, defaults to 0.1) — Beta parameter for the DPO loss.
loss_type (str, optional, defaults to "sigmoid") — Type of DPO 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 workers to use to process the data.

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

TRL

Online DPO Trainer

Overview

Usage tips

Expected dataset format

Explanation of the logged metrics

What is my model doing exactly?

Implementation details

Benchmark experiments

OnlineDPOTrainer

class trl.OnlineDPOTrainer

tokenize_row

OnlineDPOConfig

class trl.OnlineDPOConfig