baellouf/qwen3-8b-medical-QA-classification-adapter

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