DistilBERT JEE MCQ Classifier

This model is a fine-tuned DistilBERT (base uncased) designed to classify correct answers for JEE-style multiple-choice math questions. It selects the correct option among four choices (A, B, C, D).


Training Data

Source: PhysicsWallahAI JEE Main 2025 Math dataset (Jan + Apr shifts)

Filtered: Only multiple-choice questions (MCQs) were used.

Size: Combined January and April shifts, split into 80% train and 20% test.


Training Details

Base model: distilbert-base-uncased

Epochs: 10

Batch size: 4

Learning rate: 1e-5

Weight decay: 0.1


Results

Evaluation accuracy: 40%

Evaluation loss: ~1.42


Limitations

Accuracy is higher than random guess (25%) but not suitable for real exam preparation.

Trained only on Math MCQs from JEE Main 2025 dataset.

Does not handle numerical/subjective questions.


Intended Use

Research and experimentation with MCQ-style classification.

Baseline model for further fine-tuning or impro

Downloads last month
-
Safetensors
Model size
67M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Neural-Hacker/distilbert-jee-math-mcq-2025

Finetuned
(9329)
this model

Dataset used to train Neural-Hacker/distilbert-jee-math-mcq-2025