Datasets:

Modalities:
Text
Formats:
json
Size:
< 1K
ArXiv:
Libraries:
Datasets
pandas
smugri-mt-bench / README.md
adorkin's picture
Update README.md
2690ccb verified
---
configs:
- config_name: est
data_files:
- split: test
path:
- "est/est.jsonl"
- config_name: eng
data_files:
- split: test
path:
- "eng/eng.jsonl"
- config_name: liv
default: true
data_files:
- split: test
path:
- "liv/liv.jsonl"
- config_name: vro
data_files:
- split: test
path:
- "vro/vro.jsonl"
- config_name: kpv
data_files:
- split: test
path:
- "kpv/kpv.jsonl"
task_categories:
- text-generation
language:
- kpv
- vro
- liv
- et
- en
tags:
- Finno-Ugric
size_categories:
- n<1K
---
# SMUGRI-MT-Bench
This is a Finno-Ugric version of [MT-Bench](https://arxiv.org/abs/2306.05685) created to evaluate the multi-turn conversation and instruction-following capabilities of LLMs. It covers 4 (extremely) low-resource Finno-Ugric languages: Estonian, Livonian, Komi and Võro.
SMUGRI-MT-Bench comprises of 80 single and multi-turn questions organized into four topics: math, reasoning, writing, and general. The questions are handpicked from [LMSYS-Chat-1M](https://arxiv.org/abs/2309.11998) dataset and manually translated into Estonian, Võro, Komi and Livonian by fluent speakers of these languages.
Considering the extremely low-resource use case, the dataset is designed to include questions that <u>are challenging for language models</u> but <u>do not require expert knowledge from humans</u>. Additionally, the prompts were selected so that <u>translating them into other languages is feasible</u> in terms of both time and content. For further details, please refer to our [paper](https://arxiv.org/abs/2410.18902).
Additionally, the dataset contains 20 harmful prompts for each language to facilitate research on identifying and addressing potential vulnerabilities and biases in language models within extremely low-resource scenarios.
### Citation
```
@misc{purason2024llmsextremelylowresourcefinnougric,
title={LLMs for Extremely Low-Resource Finno-Ugric Languages},
author={Taido Purason and Hele-Andra Kuulmets and Mark Fishel},
year={2024},
eprint={2410.18902},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2410.18902},
}
```