mcq-deberta-v3-small
fine-tuned deberta-v3-small for the smart mcq solver challenge. given a science multiple choice question with five options a to e concatenated into one string, it predicts which option (label 0-4) is correct.
- base model: microsoft/deberta-v3-small
- task: 5-way sequence classification
- input format:
"<question> [SEP] A: <opt a> B: <opt b> ... E: <opt e>" - validation accuracy 0.99, kaggle leaderboard map@3 0.752
part of the dl genai diploma project (23f3000839). training notebook and the full project are in the github repo.
usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
tok = AutoTokenizer.from_pretrained("sarthak-wiz01/mcq-deberta-v3-small", use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained("sarthak-wiz01/mcq-deberta-v3-small")
text = "question [SEP] A: ... B: ... C: ... D: ... E: ..."
enc = tok(text, return_tensors="pt", truncation=True, max_length=256)
probs = torch.softmax(model(**enc).logits, dim=-1)