Hyperparameter search

Hyperparameters like learning rate, batch size, and number of epochs significantly affect training results. Trainer.hyperparameter_search() finds the best combination by running multiple trials, each with a different set of values, and returning the best one.

Each trial initializes a fresh model with model_init, samples new hyperparameters, runs a full training loop, and reports an objective to the search backend. The backend uses each objective to inform the next trial. After all trials complete, the best hyperparameters are returned in a ~trainer.utils.BestRun.

Initializing a model

Start each trial with a fresh model to avoid the previous runs’ state. model_init is called at the start of each trial and returns a new model instance, so every trial begins from the same initial weights.

from transformers import AutoModelForCausalLM

def model_init(trial):
    return AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")

trainer = Trainer(
    model_init=model_init,
    args=args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)

Don’t pass model= and model_init= together or Trainer raises an error.

Define the search space

Create a function that defines the search space. The format depends on the backend. If you don’t define a hp_space function, the default search covers learning_rate, num_train_epochs, and per_device_train_batch_size.

# install one of these hyperparam search backends
pip install optuna
pip install wandb
pip install ray[tune]

def hp_space(trial):
    return {
        "learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
        "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [16, 32, 64, 128]),
    }

Run the search

Provide an optional compute_objective function to define the optimization target. It defaults to eval_loss if present, or the sum of all metric values otherwise. Pass an explicit function to avoid relying on this fallback. The search backend optimizes the objective over n_trials runs in a given direction.

def compute_objective(metrics):
    return metrics["eval_loss"]

best_run = trainer.hyperparameter_search(
    hp_space=hp_space,
    compute_objective=compute_objective,
    n_trials=30,               # how many trials to run
    direction="minimize",      # or "maximize" for metrics like accuracy/F1
    backend="optuna",          # "optuna", "ray", or "wandb"
)

hyperparameter_search() returns a ~trainer.utils.BestRun containing the objective value and best hyperparameter combination.

best_run = trainer.hyperparameter_search(...)

best_run.objective        # 0.38  (best eval loss)
best_run.hyperparameters  # {"learning_rate": 5e-5, "num_train_epochs": 4, ...}

Apply the best hyperparameters to TrainingArguments and retrain on the full dataset.