# Deepfake Detector V12 - RAM Optimized (2 Hour Runtime)

## 🎯 Production-Grade Fine-tuned Ensemble (16K Samples, 2 Epochs)

### Built on V11, RAM-Safe Training for 2 Hour Runtime

This is **V12 RAM Optimized**, a fine-tuned version of the V11 ensemble detector with 30 real datasets, minimal synthetic generation, and 2-epoch high-quality fine-tuning optimized for RAM safety and 2 hour runtime.

## 📊 Performance

### V12 Ensemble Performance (Test Set - NEVER SEEN):
- **Test Accuracy**: 97.94%
- **Test Precision**: 0.9957
- **Test Recall**: 0.9486
- **Test F1 Score**: 0.9715

### Individual Models:
- **Model 1**: 95.95% val accuracy ✓ from V11
- **Model 2**: 97.40% val accuracy ✓ from V11
- **Model 3**: 96.25% val accuracy ✓ from V11

**Successfully loaded 3/3 models from V11!**

## ⚡ RAM Optimizations

### Training Configuration:
- **Epochs**: 2 (high-quality fine-tuning)
- **Batch Size**: 32 (RAM safe)
- **Target Samples**: 16K (reduced for RAM)
- **Pin Memory**: Enabled
- **Num Workers**: 2 (parallel loading)
- **Device**: GPU (CUDA) or CPU
- **Expected RAM**: ~5-6GB during training
- **Training Time**: ~1.5 hours

### RAM Safety Strategy:
- Reduced samples: 16K vs 30K (47% less data)
- Smaller batches: 32 vs 64 (50% less per batch)
- Same dataset diversity: All 50 datasets still used
- Per-dataset targets unchanged
- Should stay well under 12GB RAM

## 📦 Dataset Strategy

### Real Images (30 Datasets) - UNCHANGED:
- Core datasets: beans, cats_vs_dogs, tiny-imagenet, flowers, oxford-pets
- Classification: cifar10, mnist, fashion_mnist, caltech101, food101
- Specialized: stanford_dogs, gtsrb, eurosat, aircraft, sun397
- Medical/Scientific: patch_camelyon, NIH chest x-rays
- Target: ~8K real images with minimal synthetic (<1.5K if needed)

### Fake Images (20 Datasets) - UNCHANGED:
- GAN datasets: AFHQ, pokemon, wikiart, metfaces, celeba
- Style transfer: winter2summer, horse2zebra, watercolor2photo
- Diffusion models: pokemon-gpt4-captions, few-shot-universe
- Target: ~8K fake images with minimal synthetic (<1.5K if needed)

## 🎯 Key Features

1. **2 Epochs**: High-quality fine-tuning from V11 base
2. **RAM Safe**: 16K samples, batch 32
3. **Same Datasets**: All 50 datasets still used (30 real + 20 fake)
4. **Minimal Synthetic**: Only if <70% of target reached
5. **GPU Accelerated**: Optimized for both GPU and CPU
6. **Fine-tuned from V11**: Transfer learning from proven V11 architecture

## 💾 Training Details

- **Training Time**: 23.0 minutes (~0.4h)
- **Epochs per Model**: 2
- **Batch Size**: 16 (RAM optimized)
- **Target Samples**: 10,000
- **Models Loaded from V11**: 3/3
- **Real Datasets**: 31 (unchanged)
- **Fake Datasets**: 20 (unchanged)
- **Synthetic Used**: Minimal (only if needed)

## 🛡️ Anti-Memorization

### 80/10/10 Split (STRICT)
- **Training**: 80% (10,470 samples)
- **Validation**: 10% (1,308 samples)
- **Test**: 10% (1,310 samples) - **NEVER SEEN**

## 📄 License

MIT License

---

**Model Version**: V12 RAM Optimized (16K Dataset, 2 Epochs)
**Base Model**: ash12321/deepfake-detector-v11
**Release Date**: 2025-11-06
**Training Time**: ~1.5 hours
**Status**: Production Ready ✅ (RAM Safe + High-Quality Fine-tuning)