README.md

---
language:
- ar
license: apache-2.0
task_categories:
- multiple-choice
- question-answering
pretty_name: Arabic Accounting MCQ Evaluation Dataset
tags:
- accounting
- mcq
- arabic
- evaluation
- benchmark
- education
dataset_info:
  features:
  - name: id
    dtype: string
  - name: query
    dtype: string
  - name: answer
    dtype: string
  - name: text
    dtype: string
  - name: choices
    list: string
  - name: gold
    dtype: int64
  splits:
  - name: test
    num_bytes: 379104
    num_examples: 167
  download_size: 113685
  dataset_size: 379104
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# Arabic Accounting MCQ Evaluation Dataset

Validation and test splits for Arabic accounting MCQ with English letter choices.

## Dataset Structure

- **Format**: Multiple choice questions (4 options)
- **Language**: Arabic questions with English letter choices
- **Domain**: Accounting and finance
- **Validation**: 10% of total dataset
- **Test**: 10% of total dataset

## Fields

- `id`: Unique identifier
- `query`: Full MCQ prompt
- `answer`: Correct answer letter (a, b, c, d)
- `text`: Question text
- `choices`: Options ['a', 'b', 'c', 'd']
- `gold`: Correct answer index (0-3)

## Answer Mapping

- a → gold: 0
- b → gold: 1
- c → gold: 2
- d → gold: 3

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("SahmBenchmark/arabic-accounting-mcq_eval")

# Access splits
val_data = dataset['validation']
test_data = dataset['test']

# Evaluation
correct = 0
for example in test_data:
    model_answer = model.generate(example['query'])
    if model_answer == example['answer']:
        correct += 1

accuracy = correct / len(test_data)
```

For training data, see: `SahmBenchmark/arabic-accounting-mcq_train`