README.md

---
base_model: MCG-NJU/videomae-base-finetuned-kinetics
library_name: peft
pipeline_tag: video-classification
language:
- en
tags:
- misinformation-detection
- social-media
- fakett
- peft
- lora
- transformers
- base_model:adapter:MCG-NJU/videomae-base-finetuned-kinetics
---

# VideoMAE — Video-Only Misinformation Detection on FakeTT

**Authors:** Andrei-Gabriel Radu, Ciprian-Octavian Truică, Elena-Simona Apostol

**National University of Science and Technology POLITEHNICA Bucharest**

LoRA adapter fine-tuned from `MCG-NJU/videomae-base-finetuned-kinetics` for binary **video-only misinformation classification** on FakeTT.

This model accompanies the bachelor thesis *Misinformation Detection in Social Media Videos*.


## Results

| Dataset | Modality | Macro-F1 |
|---|---|---:|
| FakeTT | Video-only | 0.7750 |

## Model

- **Base model:** `MCG-NJU/videomae-base-finetuned-kinetics`
- **Task:** Binary misinformation classification
- **Modality:** Video-only
- **Fine-tuning:** LoRA / PEFT
- **Dataset:** FakeTT
- **Number of classes:** 2
- **Primary metric:** Macro-F1

## Training

- **LoRA rank (`r`):** 8
- **LoRA alpha:** 32
- **LoRA dropout:** 0.05
- **Target modules:** `query`, `value`, `key`, `dense`
- **Bias:** none

## Usage

```python
from peft import PeftModel
from transformers import AutoImageProcessor, AutoModelForVideoClassification

repo_id = "DS4AI-UPB/videomae-misinfo-lora"
base_model_id = "MCG-NJU/videomae-base-finetuned-kinetics"

processor = AutoImageProcessor.from_pretrained(repo_id)
base_model = AutoModelForVideoClassification.from_pretrained(
    base_model_id, num_labels=2, ignore_mismatched_sizes=True
)
model = PeftModel.from_pretrained(base_model, repo_id).eval()
```

Use the same frame sampling and preprocessing procedure as during training before passing `pixel_values` to the model.

## Intended Use

Research and benchmarking of video-only misinformation detection on short social-media videos.

## Limitations

The model does not use titles, descriptions or other textual metadata. It can miss linguistic misinformation cues and can degrade under domain shift or a different frame-sampling strategy.

## Citation

```bibtex
@thesis{radu2026misinformation,
    author = {Radu, Andrei-Gabriel and Truică, Ciprian-Octavian and Apostol, Elena-Simona},
    title  = {Misinformation Detection in Social Media Videos},
    school = {National University of Science and Technology POLITEHNICA Bucharest},
    year   = {2026}
}
```