Spaces:

Param0304
/

truthlens-ml-api

Paused

truthlens-ml-api / src /train_verification.py

Deploy TruthLens ML API to Hugging Face Space

3abf967 about 2 months ago

18.9 kB

"""
train_verification.py - Claim Verification Model (Model B) Training for TruthLens AI.
Task Definition:
MODEL B: CLAIM VERIFICATION MODEL
Input: [CLAIM] claim text [SEP] evidence text
Output: SUPPORTS (0), REFUTES (1), NOT_ENOUGH_INFO (2)
Hardware Acceleration:
- Strictly utilizes NVIDIA GPU with CUDA acceleration.
- Mixed Precision FP16 enabled.
- Gradient Accumulation to maintain effective batch size = 16 within 4GB VRAM envelope.
- Peak GPU memory tracking.
"""

import os
import sys
import json
import time
import math
from typing import Dict, List, Any, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset
import pandas as pd
import numpy as np
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    get_linear_schedule_with_warmup,
)

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from check_gpu import inspect_gpu_environment
from training_utils import (
    LABEL2ID,
    ID2LABEL,
    set_seed,
    compute_class_weights,
    compute_verification_metrics,
    plot_and_save_confusion_matrix,
    log_hardware_metrics,
)


class SciFactVerificationDataset(Dataset):
    def __init__(self, df: pd.DataFrame, tokenizer, max_length: int = 256):
        self.tokenizer = tokenizer
        self.max_length = max_length
        self.claims = df["claim"].astype(str).tolist()
        self.evidences = df["evidence"].fillna("").astype(str).tolist()
        self.labels = [LABEL2ID[l] for l in df["label"]]

    def __len__(self):
        return len(self.claims)

    def __getitem__(self, idx):
        claim = self.claims[idx]
        evidence = self.evidences[idx]
        label = self.labels[idx]

        # Natural format: [CLAIM] claim text [SEP] evidence text
        claim_clean = str(claim).strip()
        evidence_clean = str(evidence).strip() if evidence and str(evidence).strip() else "None"
        input_text = f"[CLAIM] {claim_clean} [SEP] {evidence_clean}"

        encoding = self.tokenizer(
            input_text,
            max_length=self.max_length,
            padding="max_length",
            truncation=True,
            return_tensors="pt"
        )
        item = {k: v.squeeze(0) for k, v in encoding.items()}
        item["label"] = torch.tensor(label, dtype=torch.long)
        return item


def train_epoch(
    model: nn.Module,
    dataloader: DataLoader,
    optimizer: torch.optim.Optimizer,
    scheduler: Any,
    scaler: torch.amp.GradScaler,
    criterion: nn.Module,
    device: torch.device,
    amp_dtype: torch.dtype = torch.bfloat16,
    grad_accum_steps: int = 4
) -> float:
    model.train()
    total_loss = 0.0
    optimizer.zero_grad()
    use_scaler = scaler.is_enabled()
    
    for step, batch in enumerate(dataloader, start=1):
        input_ids = batch["input_ids"].to(device)
        attention_mask = batch["attention_mask"].to(device)
        labels = batch["label"].to(device)
        
        with torch.amp.autocast("cuda", dtype=amp_dtype, enabled=device.type == "cuda"):
            outputs = model(input_ids=input_ids, attention_mask=attention_mask)
            loss = criterion(outputs.logits, labels)
            loss = loss / grad_accum_steps

        if use_scaler:
            scaler.scale(loss).backward()
        else:
            loss.backward()

        total_loss += loss.item() * grad_accum_steps
        
        if step % grad_accum_steps == 0 or step == len(dataloader):
            if use_scaler:
                scaler.unscale_(optimizer)
                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
                scaler.step(optimizer)
                scaler.update()
            else:
                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
                optimizer.step()
                
            optimizer.zero_grad()
            if scheduler is not None:
                scheduler.step()
                
    return total_loss / len(dataloader)


def evaluate(
    model: nn.Module,
    dataloader: DataLoader,
    device: torch.device,
    amp_dtype: torch.dtype = torch.bfloat16
) -> Tuple[Dict[str, Any], List[int], List[int], List[List[float]]]:
    model.eval()
    all_preds = []
    all_targets = []
    all_probs = []
    
    with torch.no_grad():
        for batch in dataloader:
            input_ids = batch["input_ids"].to(device)
            attention_mask = batch["attention_mask"].to(device)
            labels = batch["label"]
            
            with torch.amp.autocast("cuda", dtype=amp_dtype, enabled=device.type == "cuda"):
                outputs = model(input_ids=input_ids, attention_mask=attention_mask)
                probs = F.softmax(outputs.logits, dim=-1).cpu().numpy()
                preds = np.argmax(probs, axis=-1)
                
            all_preds.extend(preds.tolist())
            all_targets.extend(labels.tolist())
            all_probs.extend(probs.tolist())
            
    metrics = compute_verification_metrics(all_targets, all_preds)
    return metrics, all_targets, all_preds, all_probs


def train_scifact_verification_model(
    base_dir: str,
    model_name: str = "roberta-base",
    epochs: int = 4,
    lr: float = 2e-5,
    per_device_batch_size: int = 8,
    grad_accum_steps: int = 2,
    max_length: int = 256,
    weight_decay: float = 0.01,
    warmup_ratio: float = 0.1,
    seed: int = 42
) -> Dict[str, Any]:
    print("=" * 80)
    print(" TRUTHLENS AI — CLAIM VERIFICATION TRANSFORMER TRAINING (MODEL B)")
    print("=" * 80)
    
    # 1. Inspect Hardware and Require CUDA
    set_seed(seed)
    gpu_info = inspect_gpu_environment(require_cuda=True)
    device = torch.device("cuda")
    torch.cuda.reset_peak_memory_stats(device)
    
    data_dir = os.path.join(base_dir, "data", "final", "verification")
    models_dir = os.path.join(base_dir, "models", "verification", "scifact_deberta")
    reports_dir = os.path.join(base_dir, "reports")
    os.makedirs(models_dir, exist_ok=True)
    os.makedirs(reports_dir, exist_ok=True)
    
    train_df = pd.read_csv(os.path.join(data_dir, "scifact_train.csv"))
    valid_df = pd.read_csv(os.path.join(data_dir, "scifact_valid.csv"))
    
    print(f"Dataset: SciFact Grounded Verification")
    print(f"Train samples: {len(train_df):,} | Valid samples: {len(valid_df):,}")
    print(f"Train label distribution:\n{train_df['label'].value_counts().to_dict()}")
    
    # 2. Sequence Length Verification
    train_df["evidence"] = train_df["evidence"].fillna("").astype(str)
    comb_words = train_df.apply(lambda r: len(r["claim"].split()) + len(r["evidence"].split()), axis=1)
    max_comb_words = int(comb_words.max())
    print(f"Max combined word length: {max_comb_words} words (max_length={max_length} guarantees 0.0% truncation)")
    
    # 3. Load Tokenizer & Model
    print(f"\nLoading Pretrained Transformer: {model_name}...")
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=3)
    except Exception as e:
        print(f"Notice: {model_name} load issue ({e}). Falling back to 'roberta-base'...")
        model_name = "roberta-base"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=3)
        
    model.to(device)
    
    # 4. Datasets and Loaders
    train_dataset = SciFactVerificationDataset(train_df, tokenizer, max_length=max_length)
    valid_dataset = SciFactVerificationDataset(valid_df, tokenizer, max_length=max_length)
    
    train_loader = DataLoader(train_dataset, batch_size=per_device_batch_size, shuffle=True)
    valid_loader = DataLoader(valid_dataset, batch_size=per_device_batch_size * 2, shuffle=False)
    
    # 5. Class Weights & Loss
    train_labels = [LABEL2ID[l] for l in train_df["label"]]
    class_weights = compute_class_weights(train_labels, num_classes=3)
    print(f"Class Weights (Inverse Frequency): {class_weights}")
    criterion = nn.CrossEntropyLoss(weight=torch.tensor(class_weights, device=device))
    
    # 6. Optimizer & Scheduler
    total_training_steps = math.ceil(len(train_loader) / grad_accum_steps) * epochs
    warmup_steps = int(total_training_steps * warmup_ratio)
    
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
    scheduler = get_linear_schedule_with_warmup(
        optimizer,
        num_warmup_steps=warmup_steps,
        num_training_steps=total_training_steps
    )
    amp_dtype = torch.float16
    precision_name = "FP16"
    print(f"Mixed Precision Mode: {precision_name} (amp_dtype={amp_dtype})")
    scaler = torch.amp.GradScaler("cuda", enabled=device.type == "cuda")
    
    # 7. Training Loop
    print("\nStarting Fine-Tuning...")
    start_train_time = time.time()
    best_macro_f1 = -1.0
    best_metrics = None
    best_targets, best_preds, best_probs = [], [], []
    
    for epoch in range(1, epochs + 1):
        ep_start = time.time()
        avg_loss = train_epoch(
            model=model,
            dataloader=train_loader,
            optimizer=optimizer,
            scheduler=scheduler,
            scaler=scaler,
            criterion=criterion,
            device=device,
            amp_dtype=amp_dtype,
            grad_accum_steps=grad_accum_steps
        )
        ep_time = time.time() - ep_start
        
        # Evaluate on Validation Split
        val_metrics, val_targets, val_preds, val_probs = evaluate(
            model=model,
            dataloader=valid_loader,
            device=device,
            amp_dtype=amp_dtype
        )
        print(f"Epoch {epoch}/{epochs} ({ep_time:.1f}s) | Train Loss: {avg_loss:.4f} | Val Acc: {val_metrics['accuracy']*100:.2f}% | Val Macro F1: {val_metrics['macro_f1']*100:.2f}% | Weighted F1: {val_metrics['weighted_f1']*100:.2f}%")
        
        if val_metrics["macro_f1"] > best_macro_f1:
            best_macro_f1 = val_metrics["macro_f1"]
            best_metrics = val_metrics
            best_targets, best_preds, best_probs = val_targets, val_preds, val_probs
            
            # Save Best Model Artifacts
            model.save_pretrained(models_dir)
            tokenizer.save_pretrained(models_dir)
            print(f"  -> Saved new best model checkpoint (Macro F1: {best_macro_f1*100:.2f}%)")
            
    total_training_time = time.time() - start_train_time
    peak_vram_mb = torch.cuda.max_memory_allocated(device) / (1024 ** 2)
    
    print("\n" + "=" * 80)
    print(" TRAINING COMPLETED")
    print(f"Total Time: {total_training_time:.2f}s ({total_training_time/60:.2f} min)")
    print(f"Peak GPU VRAM Allocated: {peak_vram_mb:.2f} MB")
    print(f"Best Macro F1: {best_macro_f1*100:.2f}%")
    print("=" * 80)
    
    # 8. Save Model Metadata & Configuration
    config = {
        "model_name": model_name,
        "task": "SciFact Claim Verification (Model B)",
        "dataset": "SciFact",
        "train_samples": len(train_df),
        "valid_samples": len(valid_df),
        "epochs": epochs,
        "learning_rate": lr,
        "batch_size_per_device": per_device_batch_size,
        "grad_accum_steps": grad_accum_steps,
        "effective_batch_size": per_device_batch_size * grad_accum_steps,
        "max_length": max_length,
        "class_weights": class_weights.tolist(),
        "precision": precision_name,
        "gpu_used": gpu_info.get("gpu_name"),
        "peak_vram_mb": round(peak_vram_mb, 2),
        "training_time_seconds": round(total_training_time, 2),
        "best_metrics": best_metrics,
        "label2id": LABEL2ID,
        "id2label": ID2LABEL,
    }
    with open(os.path.join(models_dir, "training_config.json"), "w", encoding="utf-8") as f:
        json.dump(config, f, indent=2)
        
    # Create README.md inside model dir
    readme_content = f"""# TruthLens AI — Claim Verification Model (Model B)
## Overview
- **Model Name:** `{model_name}` (Fine-Tuned)
- **Dataset:** SciFact Scientific Claim Verification (`scifact_train.csv` & `scifact_valid.csv`)
- **Task:** Predict veracity verdict (`SUPPORTS`, `REFUTES`, `NOT_ENOUGH_INFO`) given claim and evidence text.
- **Hardware Acceleration:** {gpu_info.get('gpu_name')} (CUDA {gpu_info.get('cuda_version')})
- **Precision:** Mixed Precision FP16
## Performance Metrics (SciFact Validation Set: {len(valid_df)} samples)
- **Accuracy:** {best_metrics['accuracy']*100:.2f}%
- **Macro F1:** {best_metrics['macro_f1']*100:.2f}%
- **Weighted F1:** {best_metrics['weighted_f1']*100:.2f}%
- **Macro Precision:** {best_metrics['macro_precision']*100:.2f}%
- **Macro Recall:** {best_metrics['macro_recall']*100:.2f}%
### Per-Class Metrics:
- **SUPPORTS:** Precision: {best_metrics['per_class']['SUPPORTS']['precision']*100:.2f}%, Recall: {best_metrics['per_class']['SUPPORTS']['recall']*100:.2f}%, F1: {best_metrics['per_class']['SUPPORTS']['f1']*100:.2f}%
- **REFUTES:** Precision: {best_metrics['per_class']['REFUTES']['precision']*100:.2f}%, Recall: {best_metrics['per_class']['REFUTES']['recall']*100:.2f}%, F1: {best_metrics['per_class']['REFUTES']['f1']*100:.2f}%
- **NOT_ENOUGH_INFO:** Precision: {best_metrics['per_class']['NOT_ENOUGH_INFO']['precision']*100:.2f}%, Recall: {best_metrics['per_class']['NOT_ENOUGH_INFO']['recall']*100:.2f}%, F1: {best_metrics['per_class']['NOT_ENOUGH_INFO']['f1']*100:.2f}%
## Training Configuration
- **Epochs:** {epochs}
- **Learning Rate:** {lr}
- **Per-Device Batch Size:** {per_device_batch_size} (Grad Accum: {grad_accum_steps} -> Effective Batch Size: {per_device_batch_size*grad_accum_steps})
- **Max Length:** {max_length} tokens
- **Peak VRAM:** {peak_vram_mb:.2f} MB
- **Training Time:** {total_training_time:.2f} seconds
"""
    with open(os.path.join(models_dir, "README.md"), "w", encoding="utf-8") as f:
        f.write(readme_content)
        
    # 9. Plot and Save Confusion Matrix
    cm_path = os.path.join(reports_dir, "confusion_matrix.png")
    plot_and_save_confusion_matrix(
        best_metrics["confusion_matrix"],
        class_names=["SUPPORTS", "REFUTES", "NOT_ENOUGH_INFO"],
        output_path=cm_path,
        title="SciFact Transformer Model B — Confusion Matrix"
    )
    
    # 10. Perform Detailed Error Analysis
    print("\nCompiling error analysis...")
    error_list = []
    for i in range(len(valid_df)):
        actual = best_targets[i]
        pred = best_preds[i]
        if actual != pred:
            conf = float(best_probs[i][pred])
            actual_str = ID2LABEL[actual]
            pred_str = ID2LABEL[pred]
            error_list.append({
                "claim": valid_df.iloc[i]["claim"],
                "evidence": valid_df.iloc[i]["evidence"],
                "actual_label": actual_str,
                "predicted_label": pred_str,
                "confidence": round(conf, 4),
            })
            
    df_errors = pd.DataFrame(error_list).sort_values("confidence", ascending=False)
    errors_csv = os.path.join(reports_dir, "verification_errors.csv")
    df_errors.to_csv(errors_csv, index=False)
    print(f"Saved {len(df_errors)} error cases to: {errors_csv}")
    
    # 11. Save Hardware Performance Log
    hw_report = os.path.join(reports_dir, "training_hardware.md")
    log_hardware_metrics(
        output_path=hw_report,
        gpu_info=gpu_info,
        training_time_sec=total_training_time,
        total_samples=len(train_df),
        epochs=epochs,
        peak_vram_mb=peak_vram_mb,
        extra_notes=f"- Model Architecture: `{model_name}`\n- Sequence Length: {max_length}\n- Effective Batch Size: {per_device_batch_size*grad_accum_steps}\n- Class Weighting: Applied\n"
    )
    
    # 12. Save Validation Results Report
    res_report = os.path.join(reports_dir, "scifact_verification_results.md")
    cm = best_metrics["confusion_matrix"]
    res_md = f"""# TruthLens AI — SciFact Claim Verification Model Results
**Model:** `{model_name}` Fine-Tuned for Evidence Verification (Model B)  
**Hardware:** {gpu_info.get('gpu_name')} (CUDA {gpu_info.get('cuda_version')})  
**Dataset:** SciFact ({len(train_df)} train pairs, {len(valid_df)} validation pairs)  
**Date:** September 2026  
**Status:** Completed  
---
## 1. Primary Metrics (Validation Set)
| Metric | Score |
| :--- | :--- |
| **Accuracy** | **{best_metrics['accuracy']*100:.2f}%** |
| **Macro F1 Score** | **{best_metrics['macro_f1']*100:.2f}%** |
| **Weighted F1 Score** | **{best_metrics['weighted_f1']*100:.2f}%** |
| **Macro Precision** | **{best_metrics['macro_precision']*100:.2f}%** |
| **Macro Recall** | **{best_metrics['macro_recall']*100:.2f}%** |
---
## 2. Per-Class Breakdown
| Class | Precision | Recall | F1 Score | Validation Count |
| :--- | :--- | :--- | :--- | :--- |
| **`SUPPORTS`** | {best_metrics['per_class']['SUPPORTS']['precision']*100:.2f}% | {best_metrics['per_class']['SUPPORTS']['recall']*100:.2f}% | **{best_metrics['per_class']['SUPPORTS']['f1']*100:.2f}%** | {best_metrics['per_class']['SUPPORTS']['support']} |
| **`REFUTES`** | {best_metrics['per_class']['REFUTES']['precision']*100:.2f}% | {best_metrics['per_class']['REFUTES']['recall']*100:.2f}% | **{best_metrics['per_class']['REFUTES']['f1']*100:.2f}%** | {best_metrics['per_class']['REFUTES']['support']} |
| **`NOT_ENOUGH_INFO`** | {best_metrics['per_class']['NOT_ENOUGH_INFO']['precision']*100:.2f}% | {best_metrics['per_class']['NOT_ENOUGH_INFO']['recall']*100:.2f}% | **{best_metrics['per_class']['NOT_ENOUGH_INFO']['f1']*100:.2f}%** | {best_metrics['per_class']['NOT_ENOUGH_INFO']['support']} |
---
## 3. Confusion Matrix
| Actual \\ Predicted | SUPPORTS | REFUTES | NOT_ENOUGH_INFO | Total |
| :--- | :--- | :--- | :--- | :--- |
| **SUPPORTS** | {cm[0][0]} | {cm[0][1]} | {cm[0][2]} | {best_metrics['per_class']['SUPPORTS']['support']} |
| **REFUTES** | {cm[1][0]} | {cm[1][1]} | {cm[1][2]} | {best_metrics['per_class']['REFUTES']['support']} |
| **NOT_ENOUGH_INFO** | {cm[2][0]} | {cm[2][1]} | {cm[2][2]} | {best_metrics['per_class']['NOT_ENOUGH_INFO']['support']} |
---
## 4. Hardware and Throughput Summary
- **Execution Time:** {total_training_time:.2f} seconds ({total_training_time/60:.2f} minutes)
- **Peak VRAM Allocated:** {peak_vram_mb:.2f} MB (within 4096 MB laptop GPU ceiling)
- **Throughput:** {(len(train_df)*epochs)/total_training_time:.2f} samples/second
"""
    with open(res_report, "w", encoding="utf-8") as f:
        f.write(res_md)
    print(f"Saved results report -> {res_report}")
    
    return {
        "best_metrics": best_metrics,
        "peak_vram_mb": peak_vram_mb,
        "training_time_sec": total_training_time,
        "model_dir": models_dir,
    }


if __name__ == "__main__":
    root_p = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
    train_scifact_verification_model(root_p)

	"""
	train_verification.py - Claim Verification Model (Model B) Training for TruthLens AI.

	Task Definition:
	MODEL B: CLAIM VERIFICATION MODEL
	Input: [CLAIM] claim text [SEP] evidence text
	Output: SUPPORTS (0), REFUTES (1), NOT_ENOUGH_INFO (2)

	Hardware Acceleration:
	- Strictly utilizes NVIDIA GPU with CUDA acceleration.
	- Mixed Precision FP16 enabled.
	- Gradient Accumulation to maintain effective batch size = 16 within 4GB VRAM envelope.
	- Peak GPU memory tracking.
	"""

	import os
	import sys
	import json
	import time
	import math
	from typing import Dict, List, Any, Tuple
	import torch
	import torch.nn as nn
	import torch.nn.functional as F
	from torch.utils.data import DataLoader, Dataset
	import pandas as pd
	import numpy as np
	from transformers import (
	AutoTokenizer,
	AutoModelForSequenceClassification,
	get_linear_schedule_with_warmup,
	)

	sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
	from check_gpu import inspect_gpu_environment
	from training_utils import (
	LABEL2ID,
	ID2LABEL,
	set_seed,
	compute_class_weights,
	compute_verification_metrics,
	plot_and_save_confusion_matrix,
	log_hardware_metrics,
	)


	class SciFactVerificationDataset(Dataset):
	def __init__(self, df: pd.DataFrame, tokenizer, max_length: int = 256):
	self.tokenizer = tokenizer
	self.max_length = max_length
	self.claims = df["claim"].astype(str).tolist()
	self.evidences = df["evidence"].fillna("").astype(str).tolist()
	self.labels = [LABEL2ID[l] for l in df["label"]]

	def __len__(self):
	return len(self.claims)

	def __getitem__(self, idx):
	claim = self.claims[idx]
	evidence = self.evidences[idx]
	label = self.labels[idx]

	# Natural format: [CLAIM] claim text [SEP] evidence text
	claim_clean = str(claim).strip()
	evidence_clean = str(evidence).strip() if evidence and str(evidence).strip() else "None"
	input_text = f"[CLAIM] {claim_clean} [SEP] {evidence_clean}"

	encoding = self.tokenizer(
	input_text,
	max_length=self.max_length,
	padding="max_length",
	truncation=True,
	return_tensors="pt"
	)
	item = {k: v.squeeze(0) for k, v in encoding.items()}
	item["label"] = torch.tensor(label, dtype=torch.long)
	return item


	def train_epoch(
	model: nn.Module,
	dataloader: DataLoader,
	optimizer: torch.optim.Optimizer,
	scheduler: Any,
	scaler: torch.amp.GradScaler,
	criterion: nn.Module,
	device: torch.device,
	amp_dtype: torch.dtype = torch.bfloat16,
	grad_accum_steps: int = 4
	) -> float:
	model.train()
	total_loss = 0.0
	optimizer.zero_grad()
	use_scaler = scaler.is_enabled()

	for step, batch in enumerate(dataloader, start=1):
	input_ids = batch["input_ids"].to(device)
	attention_mask = batch["attention_mask"].to(device)
	labels = batch["label"].to(device)

	with torch.amp.autocast("cuda", dtype=amp_dtype, enabled=device.type == "cuda"):
	outputs = model(input_ids=input_ids, attention_mask=attention_mask)
	loss = criterion(outputs.logits, labels)
	loss = loss / grad_accum_steps

	if use_scaler:
	scaler.scale(loss).backward()
	else:
	loss.backward()

	total_loss += loss.item() * grad_accum_steps

	if step % grad_accum_steps == 0 or step == len(dataloader):
	if use_scaler:
	scaler.unscale_(optimizer)
	torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
	scaler.step(optimizer)
	scaler.update()
	else:
	torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
	optimizer.step()

	optimizer.zero_grad()
	if scheduler is not None:
	scheduler.step()

	return total_loss / len(dataloader)


	def evaluate(
	model: nn.Module,
	dataloader: DataLoader,
	device: torch.device,
	amp_dtype: torch.dtype = torch.bfloat16
	) -> Tuple[Dict[str, Any], List[int], List[int], List[List[float]]]:
	model.eval()
	all_preds = []
	all_targets = []
	all_probs = []

	with torch.no_grad():
	for batch in dataloader:
	input_ids = batch["input_ids"].to(device)
	attention_mask = batch["attention_mask"].to(device)
	labels = batch["label"]

	with torch.amp.autocast("cuda", dtype=amp_dtype, enabled=device.type == "cuda"):
	outputs = model(input_ids=input_ids, attention_mask=attention_mask)
	probs = F.softmax(outputs.logits, dim=-1).cpu().numpy()
	preds = np.argmax(probs, axis=-1)

	all_preds.extend(preds.tolist())
	all_targets.extend(labels.tolist())
	all_probs.extend(probs.tolist())

	metrics = compute_verification_metrics(all_targets, all_preds)
	return metrics, all_targets, all_preds, all_probs


	def train_scifact_verification_model(
	base_dir: str,
	model_name: str = "roberta-base",
	epochs: int = 4,
	lr: float = 2e-5,
	per_device_batch_size: int = 8,
	grad_accum_steps: int = 2,
	max_length: int = 256,
	weight_decay: float = 0.01,
	warmup_ratio: float = 0.1,
	seed: int = 42
	) -> Dict[str, Any]:
	print("=" * 80)
	print(" TRUTHLENS AI — CLAIM VERIFICATION TRANSFORMER TRAINING (MODEL B)")
	print("=" * 80)

	# 1. Inspect Hardware and Require CUDA
	set_seed(seed)
	gpu_info = inspect_gpu_environment(require_cuda=True)
	device = torch.device("cuda")
	torch.cuda.reset_peak_memory_stats(device)

	data_dir = os.path.join(base_dir, "data", "final", "verification")
	models_dir = os.path.join(base_dir, "models", "verification", "scifact_deberta")
	reports_dir = os.path.join(base_dir, "reports")
	os.makedirs(models_dir, exist_ok=True)
	os.makedirs(reports_dir, exist_ok=True)

	train_df = pd.read_csv(os.path.join(data_dir, "scifact_train.csv"))
	valid_df = pd.read_csv(os.path.join(data_dir, "scifact_valid.csv"))

	print(f"Dataset: SciFact Grounded Verification")
	print(f"Train samples: {len(train_df):,} \| Valid samples: {len(valid_df):,}")
	print(f"Train label distribution:\n{train_df['label'].value_counts().to_dict()}")

	# 2. Sequence Length Verification
	train_df["evidence"] = train_df["evidence"].fillna("").astype(str)
	comb_words = train_df.apply(lambda r: len(r["claim"].split()) + len(r["evidence"].split()), axis=1)
	max_comb_words = int(comb_words.max())
	print(f"Max combined word length: {max_comb_words} words (max_length={max_length} guarantees 0.0% truncation)")

	# 3. Load Tokenizer & Model
	print(f"\nLoading Pretrained Transformer: {model_name}...")
	try:
	tokenizer = AutoTokenizer.from_pretrained(model_name)
	model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=3)
	except Exception as e:
	print(f"Notice: {model_name} load issue ({e}). Falling back to 'roberta-base'...")
	model_name = "roberta-base"
	tokenizer = AutoTokenizer.from_pretrained(model_name)
	model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=3)

	model.to(device)

	# 4. Datasets and Loaders
	train_dataset = SciFactVerificationDataset(train_df, tokenizer, max_length=max_length)
	valid_dataset = SciFactVerificationDataset(valid_df, tokenizer, max_length=max_length)

	train_loader = DataLoader(train_dataset, batch_size=per_device_batch_size, shuffle=True)
	valid_loader = DataLoader(valid_dataset, batch_size=per_device_batch_size * 2, shuffle=False)

	# 5. Class Weights & Loss
	train_labels = [LABEL2ID[l] for l in train_df["label"]]
	class_weights = compute_class_weights(train_labels, num_classes=3)
	print(f"Class Weights (Inverse Frequency): {class_weights}")
	criterion = nn.CrossEntropyLoss(weight=torch.tensor(class_weights, device=device))

	# 6. Optimizer & Scheduler
	total_training_steps = math.ceil(len(train_loader) / grad_accum_steps) * epochs
	warmup_steps = int(total_training_steps * warmup_ratio)

	optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
	scheduler = get_linear_schedule_with_warmup(
	optimizer,
	num_warmup_steps=warmup_steps,
	num_training_steps=total_training_steps
	)
	amp_dtype = torch.float16
	precision_name = "FP16"
	print(f"Mixed Precision Mode: {precision_name} (amp_dtype={amp_dtype})")
	scaler = torch.amp.GradScaler("cuda", enabled=device.type == "cuda")

	# 7. Training Loop
	print("\nStarting Fine-Tuning...")
	start_train_time = time.time()
	best_macro_f1 = -1.0
	best_metrics = None
	best_targets, best_preds, best_probs = [], [], []

	for epoch in range(1, epochs + 1):
	ep_start = time.time()
	avg_loss = train_epoch(
	model=model,
	dataloader=train_loader,
	optimizer=optimizer,
	scheduler=scheduler,
	scaler=scaler,
	criterion=criterion,
	device=device,
	amp_dtype=amp_dtype,
	grad_accum_steps=grad_accum_steps
	)
	ep_time = time.time() - ep_start

	# Evaluate on Validation Split
	val_metrics, val_targets, val_preds, val_probs = evaluate(
	model=model,
	dataloader=valid_loader,
	device=device,
	amp_dtype=amp_dtype
	)
	print(f"Epoch {epoch}/{epochs} ({ep_time:.1f}s) \| Train Loss: {avg_loss:.4f} \| Val Acc: {val_metrics['accuracy']100:.2f}% \| Val Macro F1: {val_metrics['macro_f1']100:.2f}% \| Weighted F1: {val_metrics['weighted_f1']*100:.2f}%")

	if val_metrics["macro_f1"] > best_macro_f1:
	best_macro_f1 = val_metrics["macro_f1"]
	best_metrics = val_metrics
	best_targets, best_preds, best_probs = val_targets, val_preds, val_probs

	# Save Best Model Artifacts
	model.save_pretrained(models_dir)
	tokenizer.save_pretrained(models_dir)
	print(f" -> Saved new best model checkpoint (Macro F1: {best_macro_f1*100:.2f}%)")

	total_training_time = time.time() - start_train_time
	peak_vram_mb = torch.cuda.max_memory_allocated(device) / (1024 ** 2)

	print("\n" + "=" * 80)
	print(" TRAINING COMPLETED")
	print(f"Total Time: {total_training_time:.2f}s ({total_training_time/60:.2f} min)")
	print(f"Peak GPU VRAM Allocated: {peak_vram_mb:.2f} MB")
	print(f"Best Macro F1: {best_macro_f1*100:.2f}%")
	print("=" * 80)

	# 8. Save Model Metadata & Configuration
	config = {
	"model_name": model_name,
	"task": "SciFact Claim Verification (Model B)",
	"dataset": "SciFact",
	"train_samples": len(train_df),
	"valid_samples": len(valid_df),
	"epochs": epochs,
	"learning_rate": lr,
	"batch_size_per_device": per_device_batch_size,
	"grad_accum_steps": grad_accum_steps,
	"effective_batch_size": per_device_batch_size * grad_accum_steps,
	"max_length": max_length,
	"class_weights": class_weights.tolist(),
	"precision": precision_name,
	"gpu_used": gpu_info.get("gpu_name"),
	"peak_vram_mb": round(peak_vram_mb, 2),
	"training_time_seconds": round(total_training_time, 2),
	"best_metrics": best_metrics,
	"label2id": LABEL2ID,
	"id2label": ID2LABEL,
	}
	with open(os.path.join(models_dir, "training_config.json"), "w", encoding="utf-8") as f:
	json.dump(config, f, indent=2)

	# Create README.md inside model dir
	readme_content = f"""# TruthLens AI — Claim Verification Model (Model B)

	## Overview
	- Model Name: `{model_name}` (Fine-Tuned)
	- Dataset: SciFact Scientific Claim Verification (`scifact_train.csv` & `scifact_valid.csv`)
	- Task: Predict veracity verdict (`SUPPORTS`, `REFUTES`, `NOT_ENOUGH_INFO`) given claim and evidence text.
	- Hardware Acceleration: {gpu_info.get('gpu_name')} (CUDA {gpu_info.get('cuda_version')})
	- Precision: Mixed Precision FP16

	## Performance Metrics (SciFact Validation Set: {len(valid_df)} samples)
	- Accuracy: {best_metrics['accuracy']*100:.2f}%
	- Macro F1: {best_metrics['macro_f1']*100:.2f}%
	- Weighted F1: {best_metrics['weighted_f1']*100:.2f}%
	- Macro Precision: {best_metrics['macro_precision']*100:.2f}%
	- Macro Recall: {best_metrics['macro_recall']*100:.2f}%

	### Per-Class Metrics:
	- SUPPORTS: Precision: {best_metrics['per_class']['SUPPORTS']['precision']100:.2f}%, Recall: {best_metrics['per_class']['SUPPORTS']['recall']100:.2f}%, F1: {best_metrics['per_class']['SUPPORTS']['f1']*100:.2f}%
	- REFUTES: Precision: {best_metrics['per_class']['REFUTES']['precision']100:.2f}%, Recall: {best_metrics['per_class']['REFUTES']['recall']100:.2f}%, F1: {best_metrics['per_class']['REFUTES']['f1']*100:.2f}%
	- NOT_ENOUGH_INFO: Precision: {best_metrics['per_class']['NOT_ENOUGH_INFO']['precision']100:.2f}%, Recall: {best_metrics['per_class']['NOT_ENOUGH_INFO']['recall']100:.2f}%, F1: {best_metrics['per_class']['NOT_ENOUGH_INFO']['f1']*100:.2f}%

	## Training Configuration
	- Epochs: {epochs}
	- Learning Rate: {lr}
	- Per-Device Batch Size: {per_device_batch_size} (Grad Accum: {grad_accum_steps} -> Effective Batch Size: {per_device_batch_size*grad_accum_steps})
	- Max Length: {max_length} tokens
	- Peak VRAM: {peak_vram_mb:.2f} MB
	- Training Time: {total_training_time:.2f} seconds
	"""
	with open(os.path.join(models_dir, "README.md"), "w", encoding="utf-8") as f:
	f.write(readme_content)

	# 9. Plot and Save Confusion Matrix
	cm_path = os.path.join(reports_dir, "confusion_matrix.png")
	plot_and_save_confusion_matrix(
	best_metrics["confusion_matrix"],
	class_names=["SUPPORTS", "REFUTES", "NOT_ENOUGH_INFO"],
	output_path=cm_path,
	title="SciFact Transformer Model B — Confusion Matrix"
	)

	# 10. Perform Detailed Error Analysis
	print("\nCompiling error analysis...")
	error_list = []
	for i in range(len(valid_df)):
	actual = best_targets[i]
	pred = best_preds[i]
	if actual != pred:
	conf = float(best_probs[i][pred])
	actual_str = ID2LABEL[actual]
	pred_str = ID2LABEL[pred]
	error_list.append({
	"claim": valid_df.iloc[i]["claim"],
	"evidence": valid_df.iloc[i]["evidence"],
	"actual_label": actual_str,
	"predicted_label": pred_str,
	"confidence": round(conf, 4),
	})

	df_errors = pd.DataFrame(error_list).sort_values("confidence", ascending=False)
	errors_csv = os.path.join(reports_dir, "verification_errors.csv")
	df_errors.to_csv(errors_csv, index=False)
	print(f"Saved {len(df_errors)} error cases to: {errors_csv}")

	# 11. Save Hardware Performance Log
	hw_report = os.path.join(reports_dir, "training_hardware.md")
	log_hardware_metrics(
	output_path=hw_report,
	gpu_info=gpu_info,
	training_time_sec=total_training_time,
	total_samples=len(train_df),
	epochs=epochs,
	peak_vram_mb=peak_vram_mb,
	extra_notes=f"- Model Architecture: `{model_name}`\n- Sequence Length: {max_length}\n- Effective Batch Size: {per_device_batch_size*grad_accum_steps}\n- Class Weighting: Applied\n"
	)

	# 12. Save Validation Results Report
	res_report = os.path.join(reports_dir, "scifact_verification_results.md")
	cm = best_metrics["confusion_matrix"]
	res_md = f"""# TruthLens AI — SciFact Claim Verification Model Results

	Model: `{model_name}` Fine-Tuned for Evidence Verification (Model B)
	Hardware: {gpu_info.get('gpu_name')} (CUDA {gpu_info.get('cuda_version')})
	Dataset: SciFact ({len(train_df)} train pairs, {len(valid_df)} validation pairs)
	Date: September 2026
	Status: Completed

	---

	## 1. Primary Metrics (Validation Set)

	\| Metric \| Score \|
	\| :--- \| :--- \|
	\| Accuracy \| *{best_metrics['accuracy']100:.2f}%** \|
	\| Macro F1 Score \| *{best_metrics['macro_f1']100:.2f}%** \|
	\| Weighted F1 Score \| *{best_metrics['weighted_f1']100:.2f}%** \|
	\| Macro Precision \| *{best_metrics['macro_precision']100:.2f}%** \|
	\| Macro Recall \| *{best_metrics['macro_recall']100:.2f}%** \|

	---

	## 2. Per-Class Breakdown

	\| Class \| Precision \| Recall \| F1 Score \| Validation Count \|
	\| :--- \| :--- \| :--- \| :--- \| :--- \|
	\| `SUPPORTS` \| {best_metrics['per_class']['SUPPORTS']['precision']100:.2f}% \| {best_metrics['per_class']['SUPPORTS']['recall']100:.2f}% \| *{best_metrics['per_class']['SUPPORTS']['f1']100:.2f}%** \| {best_metrics['per_class']['SUPPORTS']['support']} \|
	\| `REFUTES` \| {best_metrics['per_class']['REFUTES']['precision']100:.2f}% \| {best_metrics['per_class']['REFUTES']['recall']100:.2f}% \| *{best_metrics['per_class']['REFUTES']['f1']100:.2f}%** \| {best_metrics['per_class']['REFUTES']['support']} \|
	\| `NOT_ENOUGH_INFO` \| {best_metrics['per_class']['NOT_ENOUGH_INFO']['precision']100:.2f}% \| {best_metrics['per_class']['NOT_ENOUGH_INFO']['recall']100:.2f}% \| *{best_metrics['per_class']['NOT_ENOUGH_INFO']['f1']100:.2f}%** \| {best_metrics['per_class']['NOT_ENOUGH_INFO']['support']} \|

	---

	## 3. Confusion Matrix

	\| Actual \\ Predicted \| SUPPORTS \| REFUTES \| NOT_ENOUGH_INFO \| Total \|
	\| :--- \| :--- \| :--- \| :--- \| :--- \|
	\| SUPPORTS \| {cm[0][0]} \| {cm[0][1]} \| {cm[0][2]} \| {best_metrics['per_class']['SUPPORTS']['support']} \|
	\| REFUTES \| {cm[1][0]} \| {cm[1][1]} \| {cm[1][2]} \| {best_metrics['per_class']['REFUTES']['support']} \|
	\| NOT_ENOUGH_INFO \| {cm[2][0]} \| {cm[2][1]} \| {cm[2][2]} \| {best_metrics['per_class']['NOT_ENOUGH_INFO']['support']} \|

	---

	## 4. Hardware and Throughput Summary
	- Execution Time: {total_training_time:.2f} seconds ({total_training_time/60:.2f} minutes)
	- Peak VRAM Allocated: {peak_vram_mb:.2f} MB (within 4096 MB laptop GPU ceiling)
	- Throughput: {(len(train_df)*epochs)/total_training_time:.2f} samples/second
	"""
	with open(res_report, "w", encoding="utf-8") as f:
	f.write(res_md)
	print(f"Saved results report -> {res_report}")

	return {
	"best_metrics": best_metrics,
	"peak_vram_mb": peak_vram_mb,
	"training_time_sec": total_training_time,
	"model_dir": models_dir,
	}


	if __name__ == "__main__":
	root_p = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
	train_scifact_verification_model(root_p)