File size: 9,335 Bytes
0f73085 df6d6d3 0f73085 7dee9a2 df6d6d3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 |
---
license: mit
task_categories:
- question-answering
language:
- id
tags:
- knowledge
pretty_name: IndoMMLU
size_categories:
- 10K<n<100K
---
# IndoMMLU
<!---
[](https://github.com/internLM/OpenCompass/) [](https://github.com/EleutherAI/lm-evaluation-harness)
-->
<p align="center"> <img src="https://raw.githubusercontent.com/fajri91/eval_picts/master/IndoMMLU-Bar.png" style="width: 100%;" id="title-icon">
</p>
<p align="center"> <a href="http://www.fajrikoto.com" target="_blank">Fajri Koto</a>, <a href="https://www.linkedin.com/in/nuaisyah/" target="_blank">Nurul Aisyah</a>, <a href="https://haonan-li.github.io/" target="_blank">Haonan Li</a>, <a href="https://people.eng.unimelb.edu.au/tbaldwin/" target="_blank">Timothy Baldwin</a> </p>
<h4 align="center">
<p align="center" style="display: flex; flex-direction: row; justify-content: center; align-items: center">
📄 <a href="https://arxiv.org/abs/2310.04928" target="_blank" style="margin-right: 15px; margin-left: 10px">Paper</a> •
🏆 <a href="https://github.com/fajri91/IndoMMLU/blob/main/README_EN.md#evaluation" target="_blank" style="margin-left: 10px">Leaderboard</a> •
🤗 <a href="https://huggingface.co/datasets/indolem/indommlu" target="_blank" style="margin-left: 10px">Dataset</a>
</p>
</h4>
## Introduction
We introduce IndoMMLU, the first multi-task language understanding benchmark for Indonesian culture and languages,
which consists of questions from primary school to university entrance exams in Indonesia. By employing professional teachers,
we obtain 14,906 questions across 63 tasks and education levels, with 46\% of the questions focusing on assessing proficiency
in the Indonesian language and knowledge of nine local languages and cultures in Indonesia.
<p align="left"> <img src="https://github.com/fajri91/eval_picts/blob/master/IndoMMLU-dist.png?raw=true" style="width: 500px;" id="title-icon"> </p>
## Subjects
| Level | Subjects |
|-----------|------------------------------------|
| SD (Primary School) | Science, Social science, Civics, Indonesian Language, Balinese, Makassarese, Banjarese, Lampungic, Madurese, Sundanese, Javanese, Dayak Ngaju, Minangkabau culture, Art, Sports, Islam religion, Christian religion, Hindu religion |
| SMP (Junior High School) | Science, Social science, Civics, Indonesian Language, Balinese, Makassarese, Banjarese, Lampungic, Madurese, Sundanese, Javanese, Minangkabau culture, Art, Sports, Islam religion, Christian religion, Hindu religion |
| SMA (Senior High School) | Physics, Chemistry, Biology, Geography, Sociology, Economics, History, Civics, Indonesian Language, Balinese, Makassarese, Banjarese, Lampungic, Madurese, Sundanese, Javanese, Art, Sports, Islam religion, Christian religion, Hindu religion |
University Entrance Test | Chemistry, Biology, Geography, Sociology, Economics, History, Indonesian Language |
We categorize the collected questions into different subject areas, including: (1) STEM (Science, Technology, Engineering, and Mathematics); (2) Social Science; (3) Humanities; (4) Indonesian Language; and (5) Local Languages and Cultures.
## Examples
These questions are written in Indonesian. For local language subjects, some are written in the local languages. The English version is for illustrative purposes only.
<p align="left">
<img src="https://github.com/fajri91/eval_picts/blob/master/min_example.png?raw=true" style="width: 400px;" id="title-icon">
</p>
## Evaluation
We evaluate 24 multilingual LLMs of different sizes in zero-shot and few-shot settings. This includes [GPT-3.5 (ChatGPT)](https://chat.openai.com/), [XGLM](https://arxiv.org/abs/2112.10668), [Falcon](https://falconllm.tii.ae/), [BLOOMZ](https://huggingface.co/bigscience/bloomz), [mT0](https://huggingface.co/bigscience/bloomz), [LLaMA](https://arxiv.org/abs/2302.13971), and [Bactrian-X](https://github.com/mbzuai-nlp/bactrian-x). Prior to the question and multiple-choice options, we add a simple prompt in the Indonesian language:
```
Ini adalah soal [subject] untuk [level]. Pilihlah salah satu jawaban yang dianggap benar!
English Translation: This is a [subject] question for [level]. Please choose the correct answer!
```
#### Zero-shot Evaluation
| Model (#param) | STEM | Social Science | Humanities | Indonesian Lang. | Local L. Culture | Average |
|---------------------|------|----------|-------------|---------|----------|---------|
| Random | 21.9 | 23.4 | 23.5 | 24.4 | 26.6 | 24.4 |
| [GPT-3.5 (175B)](https://chat.openai.com/) | **54.3** | **62.5** | **64.0** | **62.2** | 39.3 | **53.2** |
| [XGLM (564M)](https://huggingface.co/facebook/xglm-564M) | 22.1 | 23.0 | 25.6 | 25.6 | 27.5 | 25.2 |
| [XGLM (1.7B)](https://huggingface.co/facebook/xglm-1.7B) | 20.9 | 23.0 | 24.6 | 24.8 | 26.6 | 24.4 |
| [XGLM (2.9B)](https://huggingface.co/facebook/xglm-2.9B) | 22.9 | 23.2 | 25.4 | 26.3 | 27.2 | 25.2 |
| [XGLM (4.5B)](https://huggingface.co/facebook/xglm-4.5B) | 21.8 | 23.1 | 25.6 | 25.8 | 27.1 | 25.0 |
| [XGLM (7.5B)](https://huggingface.co/facebook/xglm-7.5B) | 22.7 | 21.7 | 23.6 | 24.5 | 27.5 | 24.5 |
| [Falcon (7B)](https://huggingface.co/tiiuae/falcon-7b) | 22.1 | 22.9 | 25.5 | 25.7 | 27.5 | 25.1 |
| [Falcon (40B)](https://huggingface.co/tiiuae/falcon-40b) | 30.2 | 34.8 | 34.8 | 34.9 | 29.2 | 32.1 |
| [BLOOMZ (560M)](https://huggingface.co/bigscience/bloomz-560m) | 22.9 | 23.6 | 23.2 | 24.2 | 25.1 | 24.0 |
| [BLOOMZ (1.1B)](https://huggingface.co/bigscience/bloomz-1b1) | 20.4 | 21.4 | 21.1 | 23.5 | 24.7 | 22.4 |
| [BLOOMZ (1.7B)](https://huggingface.co/bigscience/bloomz-1b7) | 31.5 | 39.3 | 38.3 | 42.8 | 29.4 | 34.4 |
| [BLOOMZ (3B)](https://huggingface.co/bigscience/bloomz-3b) | 33.5 | 44.5 | 39.7 | 46.7 | 29.8 | 36.4 |
| [BLOOMZ (7.1B)](https://huggingface.co/bigscience/bloomz-7b1) | 37.1 | 46.7 | 44.0 | 49.1 | 28.2 | 38.0 |
| [mT0<sub>small</sub> (300M)](https://huggingface.co/bigscience/mt0-small) | 21.8 | 21.4 | 25.7 | 25.1 | 27.6 | 24.9 |
| [mT0<sub>base</sub> (580M)](https://huggingface.co/bigscience/mt0-base) | 22.6 | 22.6 | 25.7 | 25.6 | 26.9 | 25.0 |
| [mT0<sub>large</sub> (1.2B)](https://huggingface.co/bigscience/mt0-large) | 22.0 | 23.4 | 25.1 | 27.3 | 27.6 | 25.2 |
| [mT0<sub>xl</sub> (3.7B)](https://huggingface.co/bigscience/mt0-xl) | 31.4 | 42.9 | 41.0 | 47.8 | 35.7 | 38.2 |
| [mT0<sub>xxl</sub> (13B)](https://huggingface.co/bigscience/mt0-xxl) | 33.5 | 46.2 | 47.9 | 52.6 | **39.6** | 42.5 |
| [LLaMA (7B)](https://arxiv.org/abs/2302.13971) | 22.8 | 23.1 | 25.1 | 26.7 | 27.6 | 25.3 |
| [LLaMA (13B)](https://arxiv.org/abs/2302.13971) | 24.1 | 23.0 | 24.4 | 29.5 | 26.7 | 25.3 |
| [LLaMA (30B)](https://arxiv.org/abs/2302.13971) | 25.4 | 23.5 | 25.9 | 28.4 | 28.7 | 26.5 |
| [LLaMA (65B)](https://arxiv.org/abs/2302.13971) | 33.0 | 37.7 | 40.8 | 41.4 | 32.1 | 35.8 |
| [Bactrian-X-LLaMA (7B)](https://github.com/mbzuai-nlp/bactrian-x) | 23.3 | 24.0 | 26.0 | 26.1 | 27.5 | 25.7 |
| [Bactrian-X-LLaMA (13B)](https://github.com/mbzuai-nlp/bactrian-x) | 28.3 | 29.9 | 32.8 | 35.2 | 29.2 | 30.3 |
#### GPT-3.5 performance (% accuracy) across different education levels
<p align="left">
<img src="https://github.com/fajri91/eval_picts/blob/master/IndoMMLU-result.png?raw=true" style="width: 370px;" id="title-icon">
</p>
Red indicates that the score is below the minimum passing threshold of 65, while green signifies a score at or above this minimum. We can observe that ChatGPT mostly passes a score of 65 in Indonesian primary school exams.
#### Few-shot Evaluation
<p align="left">
<img src="https://github.com/fajri91/eval_picts/blob/master/plot_fewshot.png?raw=true" style="width: 380px;" id="title-icon">
</p>
## Data
Each question in the dataset is a multiple-choice question with up to 5 choices and only one choice as the correct answer.
We provide our dataset according to each subject in [data](data) folder. You can also access our dataset via [Hugging Face](https://huggingface.co/datasets/indolem/indommlu).
<!--
#### Quick Use
Our dataset has been added to [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) and [OpenCompass](https://github.com/InternLM/opencompass), you can evaluate your model via these open-source tools.
-->
#### Evaluation
The code for the evaluation of each model we used is in `evaluate.py`, and the code to run them is listed in `run.sh`.
## Citation
```
@inproceedings{koto-etal-2023-indommlu,
title = "Large Language Models Only Pass Primary School Exams in {I}ndonesia: A Comprehensive Test on {I}ndo{MMLU}",
author = "Fajri Koto and Nurul Aisyah and Haonan Li and Timothy Baldwin",
booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = December,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
}
```
## License
The IndoMMLU dataset is licensed under a
[Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](http://creativecommons.org/licenses/by-nc-sa/4.0/). |