detector/train_model.py

import json
import torch
from datasets import Dataset
import evaluate
from transformers import RobertaTokenizer, RobertaForSequenceClassification, Trainer, TrainingArguments

# Load dataset from JSON
with open("ai_training_dataset.json", "r", encoding="utf-8") as f:
    data = json.load(f)

# Check that all labels are integers (0, 1, or 2)
for item in data:
    item["label"] = int(item["label"])  # Ensure type is correct

# Convert to HuggingFace Dataset
dataset = Dataset.from_list(data)
dataset = dataset.train_test_split(test_size=0.2)
train_dataset = dataset["train"]
eval_dataset = dataset["test"]

# Load tokenizer
model_name = "roberta-base"
tokenizer = RobertaTokenizer.from_pretrained(model_name)

# Tokenization function
def tokenize(example):
    return tokenizer(example["text"], truncation=True, padding="max_length", max_length=512)

# Tokenize datasets
train_dataset = train_dataset.map(tokenize, batched=True)
eval_dataset = eval_dataset.map(tokenize, batched=True)

# Keep only model-required fields
train_dataset.set_format(type="torch", columns=["input_ids", "attention_mask", "label"])
eval_dataset.set_format(type="torch", columns=["input_ids", "attention_mask", "label"])

# Load model with 3 output labels
model = RobertaForSequenceClassification.from_pretrained(model_name, num_labels=3)

# Optional: define metrics
accuracy = evaluate.load("accuracy")

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = torch.argmax(torch.tensor(logits), dim=1)
    return accuracy.compute(predictions=preds, references=labels)

# Training configuration
training_args = TrainingArguments(
    output_dir="./models/roberta-detector",
    evaluation_strategy="epoch",  # MUST match save_strategy
    save_strategy="epoch",
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    num_train_epochs=3,
    logging_steps=10,
    save_total_limit=1,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    report_to="none",  # Prevents WandB issues
)

# Trainer setup
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_metrics,
)

# Train
trainer.train()

# Save model + tokenizer
model.save_pretrained("./models/roberta-detector")
tokenizer.save_pretrained("./models/roberta-detector")

print("✅ Model trained and saved.")