ValueError
wandb.errors.Error
Stack trace
ValueError: report_to must be a list of strings containing valid logging integrations, e.g. ['wandb']
File "/usr/local/lib/python3.9/site-packages/transformers/trainer.py", line 1234, in train
self._setup_wandb()
File "/usr/local/lib/python3.9/site-packages/transformers/trainer.py", line 567, in _setup_wandb
raise ValueError("report_to must be a list of strings containing valid logging integrations, e.g. ['wandb']") Why it happens
HuggingFace Trainer expects the 'report_to' argument to be a list of strings specifying logging integrations like 'wandb'. If this argument is missing, set to None, or given as a string instead of a list, the Trainer raises a ValueError. This happens because the Trainer validates 'report_to' strictly to enable proper logging.
Detection
Check your Trainer initialization for the 'report_to' parameter; ensure it is a list of strings. Add validation or logging before training to catch invalid types early.
Causes & fixes
'report_to' parameter is set as a string instead of a list
Change 'report_to="wandb"' to 'report_to=["wandb"]' in the Trainer arguments.
'report_to' parameter is missing or set to None when wandb logging is intended
Explicitly set 'report_to=["wandb"]' in the Trainer configuration to enable wandb logging.
'report_to' list contains invalid or unsupported logging integrations
Ensure all entries in 'report_to' are valid strings recognized by HuggingFace Trainer, such as 'wandb', 'tensorboard', or 'comet_ml'.
Code: broken vs fixed
from transformers import Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
# Incorrect: report_to as string causes error
report_to="wandb" # This line triggers the ValueError
)
trainer.train()import os
from transformers import Trainer, TrainingArguments
os.environ["WANDB_API_KEY"] = os.environ.get("WANDB_API_KEY") # Ensure wandb API key is set
training_args = TrainingArguments(
output_dir="./results",
num_train_epochs=3,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
evaluation_strategy="epoch",
report_to=["wandb"] # Fixed: report_to is a list
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset
)
trainer.train()
print("Training completed with wandb logging enabled.")Workaround
If you cannot fix the 'report_to' parameter immediately, disable wandb logging by setting 'report_to=[]' or 'report_to=None' temporarily to avoid the error.
Prevention
Always specify 'report_to' as a list of valid logging integrations in Trainer arguments and validate config parameters before training to prevent wandb integration errors.