---
language: en
license: apache-2.0
base_model: bert-base-uncased
tags:
  - question-answering
  - bert
  - squad
  - extractive-qa
datasets:
  - rajpurkar/squad
metrics:
  - exact_match
  - f1
---

# BERT SQuAD Question Answering Model

A fine-tuned version of `bert-base-uncased` on [SQuAD v1.1](https://huggingface.co/datasets/rajpurkar/squad)
for **extractive question answering**.

This model finds answer spans directly within a provided context paragraph.
It does not generate new text — the answer must exist in the context.

## Model Performance
Evaluated on 1000 examples from the SQuAD v1.1 validation set:

| Metric | Score |
|---|---|
| Exact Match (EM) | 61.20 |
| F1 Score | 76.25 |

## How to Use

```python
from transformers import pipeline

qa = pipeline("question-answering", model="argha9177/bert-squad-qa")

result = qa(
    question="What is the capital of France?",
    context="France is a country in Western Europe. Its capital city is Paris."
)
print(result)
# {'answer': 'Paris', 'score': 0.98, 'start': 58, 'end': 63}
```

## Input Format
- **question**: The question to answer (string)
- **context**: The paragraph containing the answer (string)
- The answer must exist verbatim within the context
- Max combined input length: 384 tokens
- Longer contexts are handled automatically via sliding window (stride=128)

## Training Details
| Parameter | Value |
|---|---|
| Base model | bert-base-uncased |
| Dataset | rajpurkar/squad (v1.1) |
| Training samples | 8000 |
| Epochs | 2 |
| Batch size | 16 |
| Learning rate | 3e-05 |
| Max length | 384 |
| Doc stride | 128 |
| Warmup ratio | 0.1 |
| Optimizer | AdamW with linear LR decay |