technoob05/AbjadGenEval-Arabic-AI-Detection_HCMUS_TheFangs

TeX

0

2 commits

updated Jan 14, 2026

See the code

README

AbjadGenEval Arabic AI Detection πŸš€

πŸ† 1st Place Solution - HCMUS_TheFangs

Authors: [Author Names] (HCMUS)
Competition: AbjadGenEval Shared Task 1 @ ArabicNLP 2025
Result: 1st Place (F1-Score: 0.93)

Architecture (Note: Refer to eacl2023.pdf for the vector TikZ diagram)


πŸ“„ Paper

Our approach is detailed in our paper:
HCMUS_TheFangs at AbjadGenEval: Weighted Layer Pooling with Attention Fusion for Arabic AI-Generated Text Detection

Key Features

  • Weighted Layer Pooling: Dynamically learns to weigh DeBERTa-v3 layers (4-12) to capture both morphological (lower) and semantic (higher) signals.
  • Attention Pooling: Highlights specific local "watermarks" or artifacts in the sequence.
  • Performance: Significantly outperforms baselines (AraBERT) and sophisticated training techniques (Adversarial, Contrastive).

πŸ› οΈ Installation

pip install -r requirements.txt

πŸƒβ€β™‚οΈ Usage

1. Training

Run the best performing experiment (exp19):

python exp/exp19_weighted_layer_pooling.py

2. Other Experiments

We provide code for all 15+ approaches explored in our ablation study, including:

  • exp/exp04_discourse_motif.py: Discourse marker analysis
  • exp/exp05_dediacritization.py: Diacritic-aware models
  • exp/exp17_multi_contrastive.py: Contrastive Learning

πŸ“Š Leaderboard Results

RankTeamF1 ScoreAccuracyPrecisionRecall
πŸ₯‡HCMUS_TheFangs (Ours)0.930.930.970.89
πŸ₯ˆchisboizhoigay0.900.910.950.86
πŸ₯‰alizain1570.890.890.870.90

πŸ“‚ Project Structure

β”œβ”€β”€ exp/                  # Experiment scripts (PyTorch/Transformers)
β”‚   β”œβ”€β”€ exp19_weighted... # βœ… WINNING SOLUTION
β”‚   └── ...               # Ablation studies
β”œβ”€β”€ EACL_2023.../         # LaTeX Source & PDF of the Paper
β”œβ”€β”€ training.csv          # (Ignored) Training Data
└── README.md             # This file

βš–οΈ License

MIT

technoob05/AbjadGenEval-Arabic-AI-Detection_HCMUS_TheFangs

TeX

0

2 commits

updated Jan 14, 2026

See the code

README

AbjadGenEval Arabic AI Detection πŸš€

πŸ† 1st Place Solution - HCMUS_TheFangs

Authors: [Author Names] (HCMUS)
Competition: AbjadGenEval Shared Task 1 @ ArabicNLP 2025
Result: 1st Place (F1-Score: 0.93)

Architecture (Note: Refer to eacl2023.pdf for the vector TikZ diagram)


πŸ“„ Paper

Our approach is detailed in our paper:
HCMUS_TheFangs at AbjadGenEval: Weighted Layer Pooling with Attention Fusion for Arabic AI-Generated Text Detection

Key Features

  • Weighted Layer Pooling: Dynamically learns to weigh DeBERTa-v3 layers (4-12) to capture both morphological (lower) and semantic (higher) signals.
  • Attention Pooling: Highlights specific local "watermarks" or artifacts in the sequence.
  • Performance: Significantly outperforms baselines (AraBERT) and sophisticated training techniques (Adversarial, Contrastive).

πŸ› οΈ Installation

pip install -r requirements.txt

πŸƒβ€β™‚οΈ Usage

1. Training

Run the best performing experiment (exp19):

python exp/exp19_weighted_layer_pooling.py

2. Other Experiments

We provide code for all 15+ approaches explored in our ablation study, including:

  • exp/exp04_discourse_motif.py: Discourse marker analysis
  • exp/exp05_dediacritization.py: Diacritic-aware models
  • exp/exp17_multi_contrastive.py: Contrastive Learning

πŸ“Š Leaderboard Results

RankTeamF1 ScoreAccuracyPrecisionRecall
πŸ₯‡HCMUS_TheFangs (Ours)0.930.930.970.89
πŸ₯ˆchisboizhoigay0.900.910.950.86
πŸ₯‰alizain1570.890.890.870.90

πŸ“‚ Project Structure

β”œβ”€β”€ exp/                  # Experiment scripts (PyTorch/Transformers)
β”‚   β”œβ”€β”€ exp19_weighted... # βœ… WINNING SOLUTION
β”‚   └── ...               # Ablation studies
β”œβ”€β”€ EACL_2023.../         # LaTeX Source & PDF of the Paper
β”œβ”€β”€ training.csv          # (Ignored) Training Data
└── README.md             # This file

βš–οΈ License

MIT