huggingface config_error ai_generated true

ValueError: Early stopping requires `load_best_model_at_end=True` in the training arguments, but it is set to False.

ID: huggingface/trainer-early-stopping-config

Also available as: JSON · Markdown · 中文

95%Fix Rate

90%Confidence

1Evidence

2024-04-05First Seen

Version Compatibility

Version	Status	Introduced	Deprecated	Notes
transformers 4.39.0	active	—	—	—
Python 3.10	active	—	—	—

Root Cause

The `EarlyStoppingCallback` is used with `TrainingArguments` where `load_best_model_at_end` is False, but early stopping requires this flag to be True to save the best model.

generic

中文

在 `TrainingArguments` 中 `load_best_model_at_end` 为 False 时使用了 `EarlyStoppingCallback`，但提前停止需要此标志为 True 才能保存最佳模型。

Official Documentation

https://huggingface.co/docs/transformers/en/main_classes/callback#transformers.EarlyStoppingCallback

Workarounds

95% success Set `load_best_model_at_end=True` in `TrainingArguments` and also set `metric_for_best_model` and `greater_is_better` appropriately. For example: from transformers import TrainingArguments, Trainer, EarlyStoppingCallback training_args = TrainingArguments( output_dir='./results', load_best_model_at_end=True, metric_for_best_model='eval_loss', greater_is_better=False, evaluation_strategy='steps', save_strategy='steps', ) trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset, callbacks=[EarlyStoppingCallback(early_stopping_patience=3)], )
```
Set `load_best_model_at_end=True` in `TrainingArguments` and also set `metric_for_best_model` and `greater_is_better` appropriately. For example:

from transformers import TrainingArguments, Trainer, EarlyStoppingCallback
training_args = TrainingArguments(
    output_dir='./results',
    load_best_model_at_end=True,
    metric_for_best_model='eval_loss',
    greater_is_better=False,
    evaluation_strategy='steps',
    save_strategy='steps',
)
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
)
```
90% success If you don't need early stopping, remove the `EarlyStoppingCallback` from the callbacks list and rely on manual checkpoint selection.
```
If you don't need early stopping, remove the `EarlyStoppingCallback` from the callbacks list and rely on manual checkpoint selection.
```

中文步骤

在 `TrainingArguments` 中设置 `load_best_model_at_end=True`，并适当设置 `metric_for_best_model` 和 `greater_is_better`。例如：

from transformers import TrainingArguments, Trainer, EarlyStoppingCallback
training_args = TrainingArguments(
    output_dir='./results',
    load_best_model_at_end=True,
    metric_for_best_model='eval_loss',
    greater_is_better=False,
    evaluation_strategy='steps',
    save_strategy='steps',
)
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
)

如果不需要提前停止，从回调列表中移除 `EarlyStoppingCallback`，并依赖手动检查点选择。

Dead Ends

Common approaches that don't work:

90% fail
The default is False, so removing it doesn't change behavior; the error persists.
95% fail
This directly contradicts the requirement; early stopping cannot function without loading the best model.