Qwen3-8B Arabic Medical QA Classification Adapter
A QLoRA adapter for classifying Arabic medical question-answer pairs into 82 medical categories. Developed for the AbjadMed shared task at AbjadNLP 2026.
Model Details
- Model type: QLoRA adapter for sequence classification
- Language(s): Arabic
- License: Apache 2.0
- Finetuned from: Qwen/Qwen3-8B
Key Features
- All-Linear LoRA: Following "LoRA Without Regret", adapters are applied to all linear layers (not just attention), improving performance by 4.0 F1 points over attention-only LoRA.
- Dice+CE Loss: Hybrid loss function designed for Macro F1 optimization on imbalanced data.
Training Details
Training Data
AbjadMed shared task training set: 27,951 Arabic medical QA pairs across 82 categories.
Training Hyperparameters
- LoRA rank: 256
- LoRA alpha: 16
- Epochs: 3
- Precision: bfloat16
