Reward Modeling

TRL supports custom reward modeling for anyone to perform reward modeling on their dataset and model.

Check out a complete flexible example at examples/scripts/reward_modeling.py.

Expected dataset format

The RewardTrainer expects a very specific format for the dataset since the model will be trained on pairs of examples to predict which of the two is preferred. We provide an example from the Anthropic/hh-rlhf dataset below:

Therefore the final dataset object should contain two 4 entries at least if you use the default RewardDataCollatorWithPadding data collator. The entries should be named:

input_ids_chosen
attention_mask_chosen
input_ids_rejected
attention_mask_rejected

Using the RewardTrainer

After preparing your dataset, you can use the RewardTrainer in the same way as the Trainer class from 🤗 Transformers. You should pass an AutoModelForSequenceClassification model to the RewardTrainer, along with a RewardConfig which configures the hyperparameters of the training.

Leveraging 🤗 PEFT to train a reward model

Just pass a peft_config in the keyword arguments of RewardTrainer, and the trainer should automatically take care of converting the model into a PEFT model!

from peft import LoraConfig, TaskType
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from trl import RewardTrainer, RewardConfig

model = AutoModelForSequenceClassification.from_pretrained("gpt2")
peft_config = LoraConfig(
    task_type=TaskType.SEQ_CLS,
    inference_mode=False,
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
)

...

trainer = RewardTrainer(
    model=model,
    args=training_args,
    tokenizer=tokenizer,
    train_dataset=dataset,
    peft_config=peft_config,
)

trainer.train()

Adding a margin to the loss

As in the Llama 2 paper, you can add a margin to the loss by adding a margin column to the dataset. The reward collator will automatically pass it through and the loss will be computed accordingly.

def add_margin(row):
    # Assume you have a score_chosen and score_rejected columns that you want to use to compute the margin
    return {'margin': row['score_chosen'] - row['score_rejected']}

dataset = dataset.map(add_margin)

Centering rewards

In many scenarios, it’s preferable to ensure that a reward model’s output is mean zero. This is often done by first calculating the model’s average score and then subtracting it.

[Eisenstein et al., 2023] proposed an auxiliary loss function designed to directly learn a centered reward model. This auxiliary loss minimizes the squared sum of the rewards, encouraging the model to naturally produce mean-zero outputs: $\Big( R(p, r_1) + R(p, r_2) \Big)^2$

This auxiliary loss is combined with the main loss function, weighted by the parameter center_rewards_coefficient in the [RewardConfig]. By default, this feature is deactivated (center_rewards_coefficient = None).

reward_config = RewardConfig(
    center_rewards_coefficient=0.01,
    ...
)

For reference results, please refer PR #1932.

RewardTrainer

class trl.RewardTrainer

< source >

( model: Union = None args: Optional = None data_collator: Optional = None train_dataset: Optional = 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 max_length: Optional = None peft_config: Optional = None )

The RewardTrainer can be used to train your custom Reward Model. It is a subclass of the transformers.Trainer class and inherits all of its attributes and methods. It is recommended to use an AutoModelForSequenceClassification as the reward model. The reward model should be trained on a dataset of paired examples, where each example is a tuple of two sequences. The reward model should be trained to predict which example in the pair is more relevant to the task at hand.

The reward trainer expects a very specific format for the dataset. The dataset should contain two 4 entries at least if you don’t use the default RewardDataCollatorWithPadding data collator. The entries should be named

input_ids_chosen
attention_mask_chosen
input_ids_rejected
attention_mask_rejected

Optionally, you can also pass a margin entry to the dataset. This entry should contain the margin used to modulate the loss of the reward model as outlined in https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/. If you don’t pass a margin, no margin will be used.

visualize_samples

< source >

( num_print_samples: int )

Parameters

num_print_samples (int, defaults to 4) — The number of samples to print. Set to -1 to print all samples.

Visualize the reward model logits prediction

RewardConfig

class trl.RewardConfig

< 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 max_length: Optional = None dataset_num_proc: Optional = None center_rewards_coefficient: Optional = None )

Parameters

max_length (Optional[int], optional, defaults to None) — Maximum length of the sequences (prompt + completion) in the batch. This argument is required if you want to use the default data collator.
dataset_num_proc (int, optional, defaults to None) — Number of processes to use for processing the dataset.
center_rewards_coefficient (float, optional, defaults to None) — Coefficient to incentivize the reward model to output mean-zero rewards (proposed by https://huggingface.co/papers/2312.09244, Eq. 2). Recommended value: 0.01.

Configuration class for the RewardTrainer.

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

< > Update on GitHub