---
tags:
  - text-classification
  - mcq
  - cnn
license: mit
---

# CNN MCQ Model

A custom TextCNN architecture trained for MCQ answer prediction.

## Architecture
- Embedding layer: vocab_size=30522, embed_dim=140
- 3 parallel Conv1d branches (kernel sizes 3, 4, 5), 100 filters each
- Fully connected output layer (300 -> 1)

## Performance (validation)
- Accuracy / MAP@3 reported during training: **MAP@3 = 0.98125**
- Note: in the same experiment, a fine-tuned DistilBERT model scored higher (MAP@3 = 0.99375) and was selected as the primary/best model.

## Usage
```python
from modeling_cnn import TextCNN

model = TextCNN.from_pretrained("your-username/cnn-mcq-model")
model.eval()

# input_ids: tokenized input, shape (batch, seq_len)
logits = model(input_ids)
```

## Files
- `modeling_cnn.py` — model architecture definition
- `config.json` — model hyperparameters
- `model.safetensors` — trained weights