Hugging Face's logo

Datasets:
UMxYTLAILabs
/
MalayMMLU

Tasks:
Question Answering
Languages:
Malay
Size:
10K<n<100K
Tags:
knowledge
Dataset card Viewer Files Community
MalayMMLU / README.md
soonchang's picture
soonchang
Update README.md
0ba539f verified
preview code
|
raw
history blame
17.2 kB
---
task_categories:
- question-answering
language:
- ms
tags:
- knowledge
pretty_name: MalayMMLU
size_categories:
- 10K<n<100K
---
# MalayMMLU
Released on September 27, 2024
<h4 align="center">
<p>
<p align="center" style="display: flex; flex-direction: row; justify-content: center; align-items: center">
📄 <a href="https://openreview.net/pdf?id=VAXwQqkp5e" target="_blank" style="margin-right: 15px; margin-left: 10px">Paper</a> •
🤗 <a href="https://github.com/UMxYTL-AI-Labs/MalayMMLU" target="_blank" style="margin-left: 10px">Github</a>
</p>
</h4>
## Introduction
MalayMMLU is the first multitask language understanding (MLU) for Malay Language. The benchmark comprises 24,213 questions spanning both primary (Year 1-6) and secondary (Form 1-5) education levels in Malaysia, encompassing 5 broad topics that further divide into 22 subjects.
<p align="center">
<img src="imgs/MalayMMLU.png" width="250" >
</p>
| **Category** | **Subjects** |
|----------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| **STEM** | Computer Science (Secondary), Biology (Secondary), Chemistry (Secondary), Computer Literacy (Secondary), Mathematics (Primary, Secondary), Additional Mathematics (Secondary), Design and Technology (Primary, Secondary), Core Science (Primary, Secondary), Information and Communication Technology (Primary), Automotive Technology (Secondary) |
| **Language** | Malay Language (Primary, Secondary) |
| **Social science** | Geography (Secondary), Local Studies (Primary), History (Primary, Secondary) |
| **Others** | Life Skills (Primary, Secondary), Principles of Accounting (Secondary), Economics (Secondary), Business (Secondary), Agriculture (Secondary) |
| **Humanities** | Quran and Sunnah (Secondary), Islam (Primary, Secondary), Sports Science Knowledge (Secondary) |
## Result
#### Zero-shot results of LLMs on MalayMMLU (First token accuracy)
<table>
<thead>
<tr>
<th rowspan="2">Organization</th>
<th rowspan="2">Model</th>
<th rowspan="2">Vision</th>
<th colspan="7">Acc.</th>
</tr>
<tr>
<th>Language</th>
<th>Humanities</th>
<th>STEM</th>
<th>Social Science</th>
<th>Others</th>
<th>Average</th>
</tr>
</thead>
<tbody>
<tr>
<td></td>
<td>Random</td>
<td></td>
<td>38.01</td>
<td>42.09</td>
<td>36.31</td>
<td>36.01</td>
<td>38.07</td>
<td>38.02</td>
</tr>
<tr>
<td rowspan="4">OpenAI</td>
<td>GPT-4o</td>
<td style="color: green;">✔</td>
<td><strong>87.12</strong></td>
<td><strong>88.12</strong></td>
<td><strong>83.83</strong></td>
<td><strong>82.58</strong></td>
<td><strong>83.09</strong></td>
<td><strong>84.98</strong></td>
</tr>
<tr>
<td>GPT-4</td>
<td style="color: green;">✔</td>
<td><ins>82.90</ins></td>
<td><ins>83.91</ins></td>
<td>78.80</td>
<td><ins>77.29</ins></td>
<td><ins>77.33</ins></td>
<td><ins>80.11</ins></td>
</tr>
<tr>
<td>GPT-4o mini</td>
<td style="color: green;">✔</td>
<td>82.03</td>
<td>81.50</td>
<td>78.51</td>
<td>75.67</td>
<td>76.30</td>
<td>78.78</td>
</tr>
<tr>
<td>GPT-3.5</td>
<td></td>
<td>69.62</td>
<td>71.01</td>
<td>67.17</td>
<td>66.70</td>
<td>63.73</td>
<td>67.78</td>
</tr>
<tr>
<td rowspan="7">Meta</td>
<td>LLaMA-3.1 (70B)</td>
<td></td>
<td>78.75</td>
<td>82.59</td>
<td>78.96</td>
<td>77.20</td>
<td>75.32</td>
<td>78.44</td>
</tr>
<tr>
<td>LLaMA-3.1 (8B)</td>
<td></td>
<td>65.47</td>
<td>67.17</td>
<td>64.10</td>
<td>62.59</td>
<td>62.13</td>
<td>64.24</td>
</tr>
<tr>
<td>LLaMA-3 (8B)</td>
<td></td>
<td>63.93</td>
<td>66.21</td>
<td>62.26</td>
<td>62.97</td>
<td>61.38</td>
<td>63.46</td>
</tr>
<tr>
<td>LLaMA-2 (13B)</td>
<td></td>
<td>45.58</td>
<td>50.72</td>
<td>44.13</td>
<td>44.55</td>
<td>40.87</td>
<td>45.26</td>
</tr>
<tr>
<td>LLaMA-2 (7B)</td>
<td></td>
<td>47.47</td>
<td>52.74</td>
<td>48.71</td>
<td>50.72</td>
<td>48.19</td>
<td>49.61</td>
</tr>
<tr>
<td>LLaMA-3.2 (3B)</td>
<td></td>
<td>58.52</td>
<td>60.66</td>
<td>56.65</td>
<td>54.06</td>
<td>52.75</td>
<td>56.45</td>
</tr>
<tr>
<td>LLaMA-3.2 (1B)</td>
<td></td>
<td>38.88</td>
<td>43.30</td>
<td>40.65</td>
<td>40.56</td>
<td>39.55</td>
<td>40.46</td>
</tr>
<tr>
<td rowspan="8">Qwen (Alibaba)</td>
<td>Qwen 2.5 (72B)</td>
<td></td>
<td>79.09</td>
<td>79.95</td>
<td><ins>80.88</ins></td>
<td>75.80</td>
<td>75.05</td>
<td>77.79</td>
</tr>
<tr>
<td>Qwen-2.5 (32B)</td>
<td></td>
<td>76.96</td>
<td>76.70</td>
<td>79.74</td>
<td>72.35</td>
<td>70.88</td>
<td>74.83</td>
</tr>
<tr>
<td>Qwen-2-VL (7B)</td>
<td style="color: green;">✔</td>
<td>68.16</td>
<td>63.62</td>
<td>67.58</td>
<td>60.38</td>
<td>59.08</td>
<td>63.49</td>
</tr>
<tr>
<td>Qwen-2-VL (2B)</td>
<td style="color: green;">✔</td>
<td>58.22</td>
<td>55.56</td>
<td>57.51</td>
<td>53.67</td>
<td>55.10</td>
<td>55.83</td>
</tr>
<tr>
<td>Qwen-1.5 (14B)</td>
<td></td>
<td>64.47</td>
<td>60.64</td>
<td>61.97</td>
<td>57.66</td>
<td>58.05</td>
<td>60.47</td>
</tr>
<tr>
<td>Qwen-1.5 (7B)</td>
<td></td>
<td>60.13</td>
<td>59.14</td>
<td>58.62</td>
<td>54.26</td>
<td>54.67</td>
<td>57.18</td>
</tr>
<tr>
<td>Qwen-1.5 (4B)</td>
<td></td>
<td>48.39</td>
<td>52.01</td>
<td>51.37</td>
<td>50.00</td>
<td>49.10</td>
<td>49.93</td>
</tr>
<tr>
<td>Qwen-1.5 (1.8B)</td>
<td></td>
<td>42.70</td>
<td>43.37</td>
<td>43.68</td>
<td>43.12</td>
<td>44.42</td>
<td>43.34</td>
</tr>
<tr>
<td rowspan="5">Zhipu</td>
<td>GLM-4-Plus</td>
<td></td>
<td>78.04</td>
<td>75.63</td>
<td>77.49</td>
<td>74.07</td>
<td>72.66</td>
<td>75.48</td>
</tr>
<tr>
<td>GLM-4-Air</td>
<td></td>
<td>67.88</td>
<td>69.56</td>
<td>70.20</td>
<td>66.06</td>
<td>66.18</td>
<td>67.60</td>
</tr>
<tr>
<td>GLM-4-Flash</td>
<td></td>
<td>63.52</td>
<td>65.69</td>
<td>66.31</td>
<td>63.21</td>
<td>63.59</td>
<td>64.12</td>
</tr>
<tr>
<td>GLM-4</td>
<td></td>
<td>63.39</td>
<td>56.72</td>
<td>54.40</td>
<td>57.24</td>
<td>55.00</td>
<td>58.07</td>
</tr>
<tr>
<td>GLM-4<sup>††</sup> (9B)</td>
<td></td>
<td>58.51</td>
<td>60.48</td>
<td>56.32</td>
<td>55.04</td>
<td>53.97</td>
<td>56.87</td>
</tr>
<tr>
<td rowspan="3">Google</td>
<td>Gemma-2 (9B)</td>
<td></td>
<td>75.83</td>
<td>72.83</td>
<td>75.07</td>
<td>69.72</td>
<td>70.33</td>
<td>72.51</td>
</tr>
<tr>
<td>Gemma (7B)</td>
<td></td>
<td>45.53</td>
<td>50.92</td>
<td>46.13</td>
<td>47.33</td>
<td>46.27</td>
<td>47.21</td>
</tr>
<tr>
<td>Gemma (2B)</td>
<td></td>
<td>46.50</td>
<td>51.15</td>
<td>49.20</td>
<td>48.06</td>
<td>48.79</td>
<td>48.46</td>
</tr>
<tr>
<td rowspan="2">SAIL (Sea)</td>
<td>Sailor<sup>†</sup> (14B)</td>
<td></td>
<td>78.40</td>
<td>72.88</td>
<td>69.63</td>
<td>69.47</td>
<td>68.67</td>
<td>72.29</td>
</tr>
<tr>
<td>Sailor<sup>†</sup> (7B)</td>
<td></td>
<td>74.54</td>
<td>68.62</td>
<td>62.79</td>
<td>64.69</td>
<td>63.61</td>
<td>67.58</td>
</tr>
<tr>
<td>Cohere for AI</td>
<td>Command R (32B)</td>
<td></td>
<td>71.68</td>
<td>71.49</td>
<td>66.68</td>
<td>67.19</td>
<td>63.64</td>
<td>68.47</td>
</tr>
<tr>
<td>OpenGVLab</td>
<td>InternVL2 (40B)</td>
<td style="color: green;">✔</td>
<td>70.36</td>
<td>68.49</td>
<td>64.88</td>
<td>65.93</td>
<td>60.54</td>
<td>66.51</td>
</tr>
<tr>
<td>Damo (Alibaba)</td>
<td>SeaLLM-v2.5<sup>†</sup> (7B)</td>
<td></td>
<td>69.75</td>
<td>67.94</td>
<td>65.29</td>
<td>62.66</td>
<td>63.61</td>
<td>65.89</td>
</tr>
<tr>
<td rowspan="4">Mistral</td>
<td>Pixtral (12B)</td>
<td style="color: green;">✔</td>
<td>64.81</td>
<td>62.68</td>
<td>64.72</td>
<td>63.93</td>
<td>59.49</td>
<td>63.25</td>
</tr>
<tr>
<td>Mistral Small (22B)</td>
<td></td>
<td>65.19</td>
<td>65.03</td>
<td>63.36</td>
<td>61.58</td>
<td>59.99</td>
<td>63.05</td>
</tr>
<tr>
<td>Mistral-v0.3 (7B)</td>
<td></td>
<td>56.97</td>
<td>59.29</td>
<td>57.14</td>
<td>58.28</td>
<td>56.56</td>
<td>57.71</td>
</tr>
<tr>
<td>Mistral-v0.2 (7B)</td>
<td></td>
<td>56.23</td>
<td>59.86</td>
<td>57.10</td>
<td>56.65</td>
<td>55.22</td>
<td>56.92</td>
</tr>
<tr>
<td rowspan="2">Microsoft</td>
<td>Phi-3 (14B)</td>
<td></td>
<td>60.07</td>
<td>58.89</td>
<td>60.91</td>
<td>58.73</td>
<td>55.24</td>
<td>58.72</td>
</tr>
<tr>
<td>Phi-3 (3.8B)</td>
<td></td>
<td>52.24</td>
<td>55.52</td>
<td>54.81</td>
<td>53.70</td>
<td>51.74</td>
<td>53.43</td>
</tr>
<tr>
<td>01.AI</td>
<td>Yi-1.5 (9B)</td>
<td></td>
<td>56.20</td>
<td>53.36</td>
<td>57.47</td>
<td>50.53</td>
<td>49.75</td>
<td>53.08</td>
</tr>
<tr>
<td rowspan="2">Stability AI</td>
<td>StableLM 2 (12B)</td>
<td></td>
<td>53.40</td>
<td>54.84</td>
<td>51.45</td>
<td>51.79</td>
<td>50.16</td>
<td>52.45</td>
</tr>
<tr>
<td>StableLM 2 (1.6B)</td>
<td></td>
<td>43.92</td>
<td>51.10</td>
<td>45.27</td>
<td>46.14</td>
<td>46.75</td>
<td>46.48</td>
</tr>
<tr>
<td>Baichuan</td>
<td>Baichuan-2 (7B)</td>
<td></td>
<td>40.41</td>
<td>47.35</td>
<td>44.37</td>
<td>46.33</td>
<td>43.54</td>
<td>44.30</td>
</tr>
<tr>
<td>Mesolitica</td>
<td>MaLLaM-v2<sup>†</sup> (5B)</td>
<td></td>
<td>42.57</td>
<td>46.44</td>
<td>42.24</td>
<td>40.82</td>
<td>38.74</td>
<td>42.08</td>
</tr>
<tr>
<td>Yellow.ai</td>
<td>Komodo<sup>†</sup> (7B)</td>
<td></td>
<td>43.62</td>
<td>45.53</td>
<td>39.34</td>
<td>39.75</td>
<td>39.48</td>
<td>41.72</td>
</tr>
</tbody>
</table>
Highest scores are <strong>bolded</strong> and second highest scores are <ins>underlined</ins>.
† denotes LLMs fine-tuned with Southeast Asia datasets.
†† denotes open-source GLM-4.
## Citation
```bibtex
@InProceedings{MalayMMLU2024,
author = {Poh, Soon Chang and Yang, Sze Jue and Tan, Jeraelyn Ming Li and Chieng, Lawrence Leroy Tze Yao and Tan, Jia Xuan and Yu, Zhenyu and Foong, Chee Mun and Chan, Chee Seng },
title = {MalayMMLU: A Multitask Benchmark for the Low-Resource Malay Language},
booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2024},
month = {November},
year = {2024},
}
```
## Feedback
Suggestions and opinions (both positive and negative) are greatly welcome. Please contact the author by sending email to `cs.chan at um.edu.my`.