fine_tune.py

from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from datasets import load_dataset
import torch
import os

# Paths
MODEL_NAME = "Gryphe/MythoMax-L2-13b"
OUTPUT_DIR = "/app/output"  # Path where the fine-tuned model will be saved
DATASET_NAME = "mteb/toxic_conversations_50k"

# Ensure output directory exists
os.makedirs(OUTPUT_DIR, exist_ok=True)

# Load tokenizer and model
print("Loading model and tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)

# Load dataset
print("Loading dataset...")
dataset = load_dataset(DATASET_NAME)

# Validate dataset columns
print("Validating dataset columns...")
required_columns = ["text", "label", "label_text"]
for column in required_columns:
    if column not in dataset["train"].column_names:
        if column == "label":
            print("Warning: Dataset missing 'label' column. Proceeding without it.")
        else:
            raise ValueError(f"Dataset missing required column: {column}")

# Tokenize the dataset
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)

print("Tokenizing dataset...")
tokenized_datasets = dataset.map(tokenize_function, batched=True)

# Prepare data for training
print("Preparing dataset for training...")
if "label_text" in tokenized_datasets["train"].column_names:
    tokenized_datasets = tokenized_datasets.remove_columns(["text", "label_text"])
else:
    tokenized_datasets = tokenized_datasets.remove_columns(["text"])

if "label" in tokenized_datasets["train"].column_names:
    tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
else:
    # If 'label' column is missing, create a dummy labels column
    tokenized_datasets = tokenized_datasets.map(lambda x: {"labels": 0})

tokenized_datasets.set_format("torch")

# Define training arguments
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    evaluation_strategy="steps",
    eval_steps=100,
    save_steps=100,
    save_total_limit=2,
    logging_steps=10,
    per_device_train_batch_size=1,  # Adjust based on available GPU memory
    gradient_accumulation_steps=16,
    num_train_epochs=3,
    warmup_steps=500,
    weight_decay=0.01,
    logging_dir=f"{OUTPUT_DIR}/logs",
    fp16=torch.cuda.is_available(),
    push_to_hub=False,  # Set to True if you want to push to Hugging Face Hub
)

# Define Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["test"],
)

# Fine-tune the model
print("Starting training...")
trainer.train()

# Save the fine-tuned model
print("Saving fine-tuned model...")
trainer.save_model(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)

print("Fine-tuning complete!")