Add README explaining repo structure

README.md

@@ -0,0 +1,41 @@
+# AI Detection Meta-Classifier Weights
+
+Stacking meta-classifier weights for multi-modal AI content detection.
+
+## What's in this repo
+
+| File | Description |
+|------|-------------|
+| `config.json` | Full config with model IDs + meta-classifier weights |
+| `meta_classifier_weights.json` | Same weights (explicit name) |
+
+## Architecture
+
+Each modality uses **StandardScaler + LogisticRegression** on top of multiple neural model scores:
+
+| Modality | Neural Models | Meta-Classifier Features | Accuracy |
+|----------|--------------|-------------------------|----------|
+| **Image** | 4x ViT classifiers | 4 model scores + FFT slope + HF ratio (6 features) | **99.1%** |
+| **Audio** | 2x wav2vec2 classifiers | 2 model scores + 5 spectral features (7 features) | **82.0%** |
+| **Text** | Falcon-7B (Binoculars) + RoBERTa | 2 model scores + 5 statistical features (7 features) | **97.5%** |
+
+## Neural Network Models (hosted on their original HF repos)
+
+**Image:**
+- `NYUAD-ComNets/NYUAD_AI-generated_images_detector`
+- `Organika/sdxl-detector`
+- `umm-maybe/AI-image-detector`
+- `dima806/ai_vs_real_image_detection`
+
+**Audio:**
+- `DavidCombei/wav2vec2-xls-r-1b-DeepFake-AI4TRUST`
+- `Gustking/wav2vec2-large-xlsr-deepfake-audio-classification`
+
+**Text:**
+- `tiiuae/falcon-7b` + `tiiuae/falcon-7b-instruct` (Binoculars method)
+- `Hello-SimpleAI/chatgpt-detector-roberta`
+
+## Usage
+
+The weights in this repo are the **meta-classifier layer only** (scaler params + logistic regression coefficients).
+The large neural network weights are downloaded from their original HuggingFace repos at runtime.