Getting TypeError: TrainingArguments.__init__() got an unexpected keyword argument 'evaluation_strategy' when initializing SFTTrainer with TrainingArguments including evaluation_strategy='steps' and load_best_model_at_end=True.
Using:
transformers 4.54.1trl 0.20.0unsloth 2025.7.11
Have already resolved pip dependency conflicts after upgrading transformers and installing tyro/msgspec.
The error persists specifically when including evaluation_strategy in TrainingArguments. Code snippet for trainer initialization:
from trl import SFTTrainer
from transformers import DataCollatorForSeq2Seq, TrainingArguments # Import TrainingArguments
from google.colab import userdata # Import userdata
# Define the formatting function to work with the 'messages' column
def formatting_prompts_func(examples):
texts = []
for conversation in examples["messages"]: # 'examples["messages"]' is a list of conversations
# Ensure the conversation is a list and contains dictionaries
if not isinstance(conversation, list):
print(f"Skipping invalid conversation entry: {conversation}")
continue # Skip this entry if it's not a list
processed_conversation = []
for message in conversation:
# Ensure each message is a dictionary with 'role' and 'content'
if isinstance(message, dict) and 'role' in message and 'content' in message:
processed_conversation.append(message)
else:
print(f"Skipping invalid message format in conversation: {message}")
# Decide how to handle malformed messages - skipping for now
continue # Skip this message if it's not a valid dictionary format
if not processed_conversation:
print(f"Skipping conversation with no valid messages after filtering.")
continue # Skip if no valid messages were found
# Apply the chat template to the list of valid message dictionaries
try:
text = tokenizer.apply_chat_template(
processed_conversation,
tokenize=False,
add_generation_prompt=False
)
texts.append(text)
except Exception as e:
print(f"Error applying chat template to conversation: {processed_conversation}. Error: {e}")
continue # Skip conversation if chat template application fails
# print(text) # Uncomment to debug the formatted text
return texts # Return the list of strings directly
# Define TrainingArguments instead of SFTConfig
training_args = TrainingArguments(
per_device_train_batch_size = 8,
gradient_accumulation_steps = 8,
warmup_steps = 5,
num_train_epochs = 1, # Set this for 1 full training run.
# max_steps = 30,
learning_rate = 2e-4,
logging_steps = 1,
optim = "adamw_8bit",
weight_decay = 0.01,
lr_scheduler_type = "linear",
seed = 3407,
output_dir = "outputs",
report_to = "wandb", # Use this for WandB etc
# Add checkpointing and saving to Hugging Face
save_strategy="steps",
save_steps=250, # Save a checkpoint every 100 steps
push_to_hub=True,
hub_model_id="your_name/your_model", # Replace with your HF username and model name
hub_token=userdata.get('HF_TOKEN'), # Use the HF token from secrets
# Add save_total_limit and load_best_model_at_end
save_total_limit=3, # Keep a maximum of 3 checkpoints
load_best_model_at_end=True, # Load the best model based on evaluation at the end
# Add evaluation strategy for load_best_model_at_end
evaluation_strategy="steps",
eval_steps=250, # Evaluate every 100 steps
)
trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
# dataset_text_field is not needed when formatting_func returns list[str]
max_seq_length = max_seq_length,
data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer),
packing = False, # Can make training 5x faster for short sequences.
args = training_args, # Pass the TrainingArguments object
formatting_func = formatting_prompts_func, # Add the formatting function
).
Any guidance on resolving this compatibility issue to enable load_best_model_at_end is appreciated