NidhiHe/MCQ-solver

🧠 MCQ Solver — Multi-Architecture Ensemble

An AI-powered multiple-choice question solver using a 5-fold ensemble of three distinct neural architectures:

Architecture	Base Model	Parameters	Weight Size (per fold)
DeBERTa-v3	microsoft/deberta-v3-base	184M	~738 MB
RoBERTa	roberta-base	125M	~499 MB
Custom GRU	From-scratch BiGRU + Attention	~10M	~43 MB

How It Works

1. Enter a question and 5 answer choices
2. Select a model architecture (or use DeBERTa by default)
3. The 5-fold ensemble averages predictions across all folds
4. Ranked predictions with confidence scores are returned

Training Details

• Dataset: Smart MCQ Solver Challenge (2000 train / 500 test samples)
• Cross-validation: 5-fold StratifiedKFold
• Optimizer: AdamW with cosine scheduling + warmup
• Mixed precision: FP16 (when GPU available)
• Metric: MAP@3 (Mean Average Precision at 3)