alban-labs/Kapibara

---
license: apache-2.0
datasets:
- LDJnr/Capybara
language:
- en
- sq
tags:
- Physics
- Biology
- Math
- Chemistry
- Culture
- Logic
- Roleplay
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- question-answering
---

# Kapibara: Albanian Multi-turn Conversation Dataset

## Dataset Summary

Kapibara is a comprehensive Albanian language dataset designed for multi-turn conversations. It contains over 5,300 entries covering a wide range of topics including physics, biology, mathematics, chemistry, culture, and logic. The dataset is aimed at improving text generation and question-answering capabilities in the Albanian language.

## Supported Tasks

The dataset supports the following NLP tasks:
- Text Generation
- Question Answering

## Languages

The dataset primarily contains conversations in Albanian (sq), with some entries also including English (en) translations or references.

## Dataset Structure

### Data Instances

Each instance in the dataset represents a multi-turn conversation. Here's an example structure:

```json
{
"source": "General-Instruct",
"conversation": [
{
"input": "Kryeni një detyrë shkrimi krijues: Një person që zbulon se mund të flasë me kafshët.",
"output": "Ema gjithmonë kishte ndjerë një lidhje të thellë me kafshët, por ishte një pasdite me shi kur zbuloi dhuntinë e saj të vërtetë. ..."
},
{
"input": "Diskutoni implikimet psikologjike të aftësisë së Emës për të komunikuar me kafshët në jetën e saj personale dhe sociale.",
"output": "Aftësia e re e Emës për të komunikuar me kafshët mund të ketë implikime të thella psikologjike si në jetën e saj personale ashtu edhe në atë sociale. ..."
}
]
}
```
### Data Fields

source: The source or category of the conversation.
conversation: A list of conversation turns.

input: The input or question in the conversation.
output: The corresponding output or answer.

### Data Splits
The dataset is currently provided as a single file: rough5300entries.jsonl, containing approximately 5,300 conversation entries.
Dataset Creation
Curation Rationale
The Kapibara dataset was created to address the lack of comprehensive, multi-turn conversation datasets in the Albanian language. It aims to provide a rich resource for developing and testing language models capable of understanding and generating Albanian text across various domains.
Source Data
The conversations in this dataset were carefully curated and generated to cover a wide range of topics relevant to Albanian culture and general knowledge.
Annotations
The dataset does not contain additional annotations beyond the conversation structure.
Considerations for Using the Data
Social Impact of Dataset
This dataset aims to improve NLP capabilities in the Albanian language, potentially leading to better language technologies and applications for Albanian speakers.
Discussion of Biases
While efforts have been made to cover a diverse range of topics, users should be aware of potential biases in the dataset, including but not limited to topic selection and language style.
Other Known Limitations
The dataset is limited to text-based conversations and does not include other modalities such as images or audio.

## Citation

If you use this data in your work, please cite:

```bibtex
@article{daniel2024llm,
title={MultiLLM Mix for Data Mutation and Synthesis},
author={Nisten Tahiraj, Daniel Merja, Benjamin Shehu, Jeton Kukalaj and Amittai Groot},
journal={arXiv preprint arXiv:(comming soon)},
year={2024}
}
```