DS4AI-UPB/distilbert-videomae-misinfo

DistilBERT + VideoMAE — Multimodal Misinformation Detection on FakeTT

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

National University of Science and Technology POLITEHNICA Bucharest

Supervised multimodal misinformation classification checkpoint based on DistilBERT + VideoMAE, trained and evaluated on FakeTT.

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

Results

Dataset	Modality	Macro-F1
FakeTT	Multimodal	0.8006

Model

• Architecture: DistilBERT + VideoMAE
• Task: Binary misinformation classification
• Modality: Text + video
• Dataset: FakeTT
• Number of classes: 2
• Checkpoint: pytorch_model.bin
• Primary metric: Macro-F1

Saved Metadata

• batch_size: 2
• dropout: 0.1
• early_stopping_patience: 3
• fusion_hidden_dim: 256
• gradient_accumulation_steps: 1
• learning_rate: 2e-05
• lr_scheduler_type: linear
• max_grad_norm: 1.0
• num_labels: 2
• num_train_epochs: 30
• text_lora_alpha: 16
• text_lora_dropout: 0.05
• text_lora_r: 8
• text_max_length: 256
• text_model_name: distilbert-base-uncased
• text_proj_dim: 256
• video_lora_alpha: 32
• video_lora_dropout: 0.05
• video_lora_r: 16
• video_model_name: MCG-NJU/videomae-base-finetuned-kinetics
• video_proj_dim: 256
• warmup_ratio: 0.0
• weight_decay: 0.01

The original machine-readable metadata is included as model_metadata.json.

Usage

This repository contains a checkpoint for the custom multimodal classifier used in the thesis implementation. It is not a drop-in AutoModel.from_pretrained() repository.

Instantiate the matching custom classifier, load the included tokenizer and image/video processor, then load the state dictionary:

import torch

state = torch.load("pytorch_model.bin", map_location="cpu")

# model = MatchingClassifier(...)
# model.load_state_dict(state)
# model.eval()

The constructor and preprocessing must match the thesis implementation and model_metadata.json.

Intended Use

Research, benchmarking and reproducibility of supervised multimodal misinformation detection on short social-media videos.

Limitations

The model performs classification rather than factual verification, can inherit biases from FakeTT and its pretrained encoders, and requires the matching custom implementation for reproducible inference.

Citation

@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}
}