mcq-deberta-v3-small

ONNX (int8, per-channel dynamic quantization) export of a microsoft/deberta-v3-small fine-tuned for 5-way MCQ answer selection, for the IITM BS DA2001P Deep Learning & GenAI project (Shubham Gupta, roll 22f3001874).

Input is one flattened string, "{question}\nA. {A}\nB. {B}\nC. {C}\nD. {D}\nE. {E}", max_length=160. The five classes are the option letters A-E (LABEL_0..LABEL_4).

Fine-tuning: all 2,000 labelled rows of the smart-mcq-solver-challenge train set, 3 epochs, AdamW lr 2e-5, weight decay 0.01, batch size 8, float32, linear schedule with 50 warmup steps.

Quantization fidelity (measured on 300 labelled rows, versus the fp32 PyTorch checkpoint, for the 0.7 x DeBERTa + 0.3 x RoBERTa ensemble these two models form): top-1 agreement 99.67%, MAP@3 1.000000 -> 0.998333. Top-1 predictions are essentially preserved; the very-low-probability rank 2-3 tail can reorder, because those probabilities (~0.001) are below int8 resolution.

Used by the live demo: https://huggingface.co/spaces/unlikefraction/smart-mcq-solver Source: https://github.com/unlikefraction/iitm-da2001p-may2026