models/train_model_v2.py

import os
import pandas as pd
import torch
from sklearn.model_selection import train_test_split
from transformers import DistilBertTokenizerFast, DistilBertForSequenceClassification, Trainer, TrainingArguments
from datasets import Dataset
import numpy as np
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, confusion_matrix
from sklearn.utils.class_weight import compute_class_weight

# Load dataset
df = pd.read_csv("D:\Sentinel\data\sentinel_dataset_expanded.csv")

# Basic cleaning
df = df.dropna()
df["text"] = df["text"].astype(str)

print(f"Dataset size: {len(df)}")
print(f"Label distribution: {df['label'].value_counts().to_dict()}")
print(f"\nAverage text length: {df['text'].str.len().mean():.1f} characters")

# Train-test split
train_texts, val_texts, train_labels, val_labels = train_test_split(
    df["text"].tolist(),
    df["label"].tolist(),
    test_size=0.2,
    random_state=42,
    stratify=df["label"].tolist()  # Ensure balanced split
)

# Compute class weights for balanced training
class_weights = compute_class_weight(
    class_weight='balanced',
    classes=np.unique(train_labels),
    y=train_labels
)
class_weights = torch.tensor(class_weights, dtype=torch.float)
print(f"\nClass weights: SAFE={class_weights[0]:.4f}, SCAM={class_weights[1]:.4f}")
print(f"Training set - SAFE: {train_labels.count(0)}, SCAM: {train_labels.count(1)}")

# Tokenizer
tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")

train_encodings = tokenizer(train_texts, truncation=True, padding=True, max_length=128)
val_encodings = tokenizer(val_texts, truncation=True, padding=True, max_length=128)

train_dataset = Dataset.from_dict({
    "input_ids": train_encodings["input_ids"],
    "attention_mask": train_encodings["attention_mask"],
    "labels": train_labels
})

val_dataset = Dataset.from_dict({
    "input_ids": val_encodings["input_ids"],
    "attention_mask": val_encodings["attention_mask"],
    "labels": val_labels
})

# Model with class weights
model = DistilBertForSequenceClassification.from_pretrained(
    "distilbert-base-uncased",
    num_labels=2
)

# Custom trainer class to use class weights
class WeightedTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs.pop("labels")
        outputs = model(**inputs)
        logits = outputs.logits
        loss_fct = torch.nn.CrossEntropyLoss(weight=class_weights.to(logits.device))
        loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
        return (loss, outputs) if return_outputs else loss

# Metrics
def compute_metrics(pred):
    labels = pred.label_ids
    preds = np.argmax(pred.predictions, axis=1)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')
    acc = accuracy_score(labels, preds)
    
    # Confusion matrix
    cm = confusion_matrix(labels, preds)
    print(f"\nConfusion Matrix:")
    print(f"              Predicted SAFE  Predicted SCAM")
    print(f"Actual SAFE:  {cm[0][0]:14d}  {cm[0][1]:14d}")
    print(f"Actual SCAM:  {cm[1][0]:14d}  {cm[1][1]:14d}")
    
    return {
        "accuracy": acc,
        "f1": f1,
        "precision": precision,
        "recall": recall
    }

# Training args - MORE aggressive parameters
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=10,  # Increased to 10 epochs
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    warmup_steps=50,  # Reduced warmup
    weight_decay=0.01,
    learning_rate=3e-5,  # Slightly higher learning rate
    logging_dir="./logs",
    logging_steps=5,
    eval_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    metric_for_best_model="f1",
    save_total_limit=2
)

trainer = WeightedTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    compute_metrics=compute_metrics
)

print("\n" + "="*60)
print("Starting training with 10 epochs...")
print("="*60)
trainer.train()

print("\n" + "="*60)
print("Final Evaluation...")
print("="*60)
eval_results = trainer.evaluate()
print(f"\nFinal Evaluation Results:")
print(f"  Accuracy: {eval_results['eval_accuracy']:.4f}")
print(f"  F1 Score: {eval_results['eval_f1']:.4f}")
print(f"  Precision: {eval_results['eval_precision']:.4f}")
print(f"  Recall: {eval_results['eval_recall']:.4f}")

script_dir = os.path.dirname(os.path.abspath(__file__))
save_path = os.path.join(script_dir, "sentinel_model")
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)

print(f"\n{'='*60}")
print(f"Model training complete and saved to: {save_path}")
print("="*60)