|
|
--- |
|
|
language: |
|
|
- en |
|
|
license: apache-2.0 |
|
|
library_name: unsloth |
|
|
tags: |
|
|
- llama |
|
|
- llama-3 |
|
|
- text-generation |
|
|
- deep-learning |
|
|
- image-analysis |
|
|
- deepfake-detection |
|
|
- lora |
|
|
- fine-tuning |
|
|
datasets: |
|
|
- custom |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
# Deepfake Explanation Model based on Llama 3.2 |
|
|
|
|
|
This model is fine-tuned to provide technical and non-technical explanations of deepfake detection results. It analyzes detection metrics, activation regions, and image features to explain why an image was classified as real or fake. |
|
|
|
|
|
## Model Details |
|
|
- Base model: Llama 3.2 3B Instruct |
|
|
- Training method: LoRA fine-tuning with Unsloth |
|
|
- Training data: Custom dataset of deepfake detection results with expert explanations |
|
|
|
|
|
## Use Cases |
|
|
|
|
|
This model can be used to: |
|
|
- Generate expert-level technical explanations of deepfake detection results |
|
|
- Provide simplified, accessible explanations for non-technical audiences |
|
|
- Analyze activation regions in images to explain detection decisions |
|
|
- Support educational content about deepfake detection |
|
|
|
|
|
## Usage Example |
|
|
|
|
|
```python |
|
|
from unsloth import FastLanguageModel |
|
|
import torch |
|
|
|
|
|
# Load the model |
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name="saakshigupta/deepfake-explainer-llama32", |
|
|
max_seq_length=2048, |
|
|
load_in_4bit=True, |
|
|
) |
|
|
|
|
|
# Enable for inference |
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
# Example prompt |
|
|
prompt = """Analyze this deepfake detection result and provide both a technical expert explanation and a simple non-technical explanation. |
|
|
|
|
|
Below is a deepfake detection result with explanation metrics. Provide both a technical and accessible explanation of why this image is classified as it is. |
|
|
### Detection Results: |
|
|
Verdict: Deepfake |
|
|
Confidence: 0.87 |
|
|
### Analysis Metrics: |
|
|
High Activation Regions: lips, nose |
|
|
Medium Activation Regions: eyes, chin |
|
|
Low Activation Regions: forehead, background |
|
|
Frequency Analysis Score: 0.79 |
|
|
### Image Description: |
|
|
A man with glasses and short hair looking directly at the camera. |
|
|
### Heatmap Description: |
|
|
The heatmap shows intense red coloration around the lips and nose area, suggesting these regions contributed most to the detection verdict.""" |
|
|
|
|
|
# Format for chat |
|
|
messages = [ |
|
|
{"role": "user", "content": prompt}, |
|
|
] |
|
|
|
|
|
# Apply chat template |
|
|
inputs = tokenizer.apply_chat_template( |
|
|
messages, |
|
|
tokenize=True, |
|
|
add_generation_prompt=True, |
|
|
return_tensors="pt", |
|
|
).to("cuda" if torch.cuda.is_available() else "cpu") |
|
|
|
|
|
# Generate response |
|
|
from transformers import TextStreamer |
|
|
text_streamer = TextStreamer(tokenizer, skip_prompt=True) |
|
|
_ = model.generate( |
|
|
input_ids=inputs, |
|
|
streamer=text_streamer, |
|
|
max_new_tokens=800, |
|
|
use_cache=True, |
|
|
temperature=0.7, |
|
|
do_sample=True |
|
|
) |
