---
license: mit
library_name: pytorch
tags:
- scam-detection
- multi-modal
- audio-classification
- text-classification
- fusion
- MiniLM
- vosk
---

# MultiModal Scam Detection — Models & Dataset

**Hugging Face asset repository** for the [MultiModal Scam Detection](https://github.com/Codexx121/MultiModal_Scam_Detct) project.

This repo contains **trained model checkpoints, cached features, embeddings, and test audio** — too large for GitHub.

## Contents

| Asset | Size | Description |
|-------|------|-------------|
| `audio_features/` | ~3.3 GB | Pre-computed MFCC features (2407 `.pt` files) for audio encoder training |
| `detection_checkpoints/` | ~1.1 GB | Fine-tuned MiniLM text classifier checkpoints (4 checkpoints) |
| `fusion_embeddings/` | ~14 MB | Pre-extracted audio + text embeddings + fusion dataset (`.npz`) |
| `test_samples/` | ~22 MB | Sample WAV files for testing inference |

## Download

### Via Python

```python
from huggingface_hub import snapshot_download

snapshot_download("Codex12/MultiModal_Scam_Models-Dataset", repo_type="model")
```

### Via CLI

```bash
huggingface-cli download Codex12/MultiModal_Scam_Models-Dataset --repo-type model --local-dir ./assets
```

### Via Git LFS (advanced)

```bash
git lfs install
git clone https://huggingface.co/Codex12/MultiModal_Scam_Models-Dataset
```

## Usage

```python
from huggingface_hub import hf_hub_download
import torch

# Download a checkpoint
checkpoint = hf_hub_download(
    "Codex12/MultiModal_Scam_Models-Dataset",
    "detection_checkpoints/best_model/model.safetensors",
    repo_type="model"
)

# Download audio features
feature_path = hf_hub_download(
    "Codex12/MultiModal_Scam_Models-Dataset",
    "audio_features/legitimate_00001.pt",
    repo_type="model"
)
```

## Related

- **GitHub (code)**: [Codexx121/MultiModal_Scam_Detct](https://github.com/Codexx121/MultiModal_Scam_Detct)
- **Pipeline**: Audio → MFCC → Conv2D Encoder (128-D) + ASR → MiniLM (384-D) → Fusion MLP → SCAM/LEGITIMATE