[Python] evaluation_strategy() not supported in transformers library

I've recently been attempting to fine-tune the pretrained neural network, "cardiffnlp/twitter-roberta-base-offensive." However, when I attempted to run a Python program I created, "evaluation_strategy" was considered to be an "unexpected argument" regardless that I have version 4.52.6 installed. Could anyone help me with fixing this error? (There also might be another one with the classifiers since I'm attempting to have it transition to a 6-label classifier rather than a binary classifier since the Jigsaw Dataset uses 6 labels, so if that seems incorrect in my code, please fix that as well).

Here's the source code:


from datasets import load_dataset
from transformers import *
import torch
import numpy as np
from sklearn.metrics import f1_score, accuracy_score
from sklearn.preprocessing import MultiLabelBinarizer

# Load dataset (adjust path if needed)
dataset = load_dataset('csv', data_files={
"train": "data/train_split.csv",
"validation": "data/validation_split.csv"
})

# Define label columns used in Jigsaw multi-label setup
LABEL_COLUMNS = ["toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"]

# Tokenizer and model setup
MODEL_NAME = "cardiffnlp/twitter-roberta-base-offensive"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# Define label count and problem type
config = AutoConfig.from_pretrained(
MODEL_NAME,
num_labels=len(LABEL_COLUMNS),
problem_type="multi_label_classification"
)

model = AutoModelForSequenceClassification.from_pretrained(
MODEL_NAME,
config=config,
ignore_mismatched_sizes=True
)

# Preprocessing function
def preprocess(example):
encoding = tokenizer(
example["comment_text"],
truncation=True,
padding="max_length",
max_length=128
)
labels = [example[col] for col in LABEL_COLUMNS]
encoding["labels"] = torch.tensor(labels, dtype=torch.float)
return encoding

# Apply preprocessing
encoded_dataset = dataset.map(preprocess)

# Training configuration
training_args = TrainingArguments(
output_dir="./results",
save_strategy="epoch",
evaluation_strategy="epoch", # <-- Add this line
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=3,
weight_decay=0.01,
logging_dir='./logs',
logging_steps=10,
load_best_model_at_end=True,
metric_for_best_model="f1"
)


# Metric function
def compute_metrics(pred):
preds = torch.sigmoid(torch.tensor(pred.predictions)).numpy()
labels = pred.label_ids
preds = (preds > 0.5).astype(int)
f1 = f1_score(labels, preds, average="macro")
acc = accuracy_score(labels, preds)
return {"f1": f1, "accuracy": acc}

# Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=encoded_dataset["train"],
eval_dataset=encoded_dataset["validation"],
tokenizer=tokenizer,
compute_metrics=compute_metrics
)

# Train the model
trainer.train()


Could anyone fix it?

Continue reading...