---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: language
    dtype: string
  - name: question
    dtype: string
  - name: choices
    struct:
    - name: text
      sequence: string
    - name: label
      sequence: string
  - name: answerKey
    dtype: string
  splits:
  - name: train
    num_bytes: 357825
    num_examples: 1119
  - name: validation
    num_bytes: 98118
    num_examples: 299
  - name: test
    num_bytes: 382265
    num_examples: 1172
  download_size: 433794
  dataset_size: 838208
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: apache-2.0
language:
- sw
---

# Dataset Card for ARC_Challenge_Swahili

## Dataset Summary

ARC_Challenge_Swahili is a Swahili translation of the original English ARC (AI2 Reasoning Challenge) dataset. This dataset evaluates the ability of AI systems to answer grade-school level multiple-choice science questions. The Swahili version was created using a combination of machine translation and human annotation to ensure high-quality and accurate translations.

## Translation Methodology
The translation process for the ARC_Challenge_Swahili dataset involved two main stages:

### Machine Translation:
1. The initial translation from English to Swahili was performed using the SeamlessM4TModel translation model.

* The following parameters were used for the translation:
```python
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=1024).to(device)
outputs = model.generate(**inputs, tgt_lang=dest_lang)
translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)
```

2. Human Verification and Annotation:

* After the initial machine translation, the translations were passed through GPT-3.5 for verification. This step involved checking the quality of the translations and identifying any that were not up to standard.
* Human translators reviewed and annotated the translations flagged by GPT-3.5 as problematic to ensure accuracy and naturalness in Swahili.

## Supported Tasks and Leaderboards
* multiple-choice: The dataset supports multiple-choice question-answering tasks.

## Languages
The dataset is in Swahili.

## Dataset Structure

### Data Instances
* An example of a data instance:
```json
{
  "id": "example-id",
  "language": "sw",
  "question": "Ni gani kati ya zifuatazo ni sehemu ya mmea?",
  "choices": [
    {"text": "Majani", "Jiwe", "Ubao", "Nondo"},
    {"label": "A", "B": "C", "D"},
  ],
  "answerKey": "A"
}
```

### Data Fields
* id: Unique identifier for each question.
* language: The language of the question is Swahili (sw).
* question: The science question in Swahili.
* Choices: There are multiple-choice options, each with text and label.
* answerKey: The correct answer for each question.

## Datasplit
| Split | Num Rows |
|-------|----------|
| train | 1119 |
| validation | 299 |
| test | 1172 |