File size: 9,811 Bytes
cc4c4bb 2b10415 cc4c4bb 845c7da cc4c4bb 2b10415 cc4c4bb 845c7da cc4c4bb 2b10415 845c7da 2b10415 cc4c4bb 2b10415 cc4c4bb 2b10415 845c7da 2b10415 845c7da 2b10415 845c7da 2b10415 845c7da 2b10415 845c7da 2b10415 845c7da 2b10415 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 |
---
library_name: transformers
tags: []
---
# KO-REAson
**KO-REAson** is a series of Korean-centric reasoning language models developed in collaboration with [OneLineAI](https://onelineai.com/), [KISTI-KONI](https://huggingface.co/KISTI-KONI), [HAE-RAE](https://huggingface.co/HAERAE-HUB) and ORACLE.
We use the **Language-Mixed Chain-of-Thought (CoT)** approach, which allows the model to alternate between English and Korean during the “Think” stage of reasoning, preserving key Korean terms while leveraging English for logical scaffolding.
Top-performing models of our series [KO-REAson-AX3_1-7B-0831 (KONI-7B-R-20250831)](https://huggingface.co/KISTI-KONI/KONI-7B-R-20250831) and [KO-REAson-7B-Q2_5-0831](https://huggingface.co/KoReason/KO-REASon-7B-Q2_5-0831) show performance comparable to models trained on closed-source datasets such as Exaone-Deep-7.8B.
<p align="left">
<img src="https://cdn-uploads.huggingface.co/production/uploads/60d3e619b8448e1785bbda2a/uqrKdxbQEqAFknYBmuH7Y.png"
alt="Model Comparison" width="750"/>
<br>
<em style="display:inline-block; max-width:750px; text-align:cener; white-space:normal; word-wrap:break-word; line-height:1.5;">
<b>Left:</b> Average performance (Held-out-Ko) of open models trained on closed or open data;
our models are highlighted in green.
</em>
</p>
## Model Details
The **KO-REAson-0831** family comes in six variants based on the base model used.
| Model (link) | Base | Notes |
| -------------------------------------------------------------------------------------------- | -------------------- | --------------------------- |
| [KO-REAson-L3_1-8B-0831](https://huggingface.co/KoReason/KO-REASon-L3_1-8B-0831) | [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | `L3_1` → Llama-3.1-8B |
| [KO-REAson-KL3_1-8B-0831](https://huggingface.co/KOREAson/KO-REAson-KL3_1-8B-0831) | [Koni-Llama-3.1-8B](https://huggingface.co/KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024) | `KL3_1` → Koni-Llama-3.1-8B; also called [KONI-Llama3.1-8B-R-20250831](https://huggingface.co/KISTI-KONI/KONI-Llama3.1-8B-R-20250831) |
| [KO-REAson-G3-4B-0831](https://huggingface.co/KoReason/KO-REASon-G3-4B-0831) | [Gemma-3 4B](https://huggingface.co/google/gemma-3-4b-it) | `G3` → Gemma-3-4B |
| [KO-REAson-AX3_1-7B-0831](https://huggingface.co/KOREAson/KO-REAson-7B-AX3_1-0831) | [A.X.-3.1-Light (≈7B)](https://huggingface.co/skt/A.X-3.1-Light) | `AX3_1` → A.X.-3.1-Light; also called [KONI-7B-R-20250831](https://huggingface.co/KISTI-KONI/KONI-7B-R-20250831) |
| [KO-REAson-K2505_8B-0831](https://huggingface.co/KoReason/KO-REASon-K2505_8B-0831) | [Kanana-2505 (8B)](https://huggingface.co/kakaocorp/kanana-1.5-8b-instruct-2505) | `K2505` → Kanana-2505 |
| [KO-REAson-7B-Q2_5-0831](https://huggingface.co/KoReason/KO-REASon-7B-Q2_5-0831) | [Qwen-2.5 (7B)](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) | `Q2_5` → Qwen-2.5 |
# Performance
**Evaluation Datasets**
The model's performance was evaluated across a total of 11 benchmarks, and the evaluation suite is divided into two parts: (You can check these benchmarks in [HAERAE-HUB/KoSimpleEval](https://huggingface.co/datasets/HAERAE-HUB/KoSimpleEval))
- **Held-in**: This set of benchmarks is used for routine monitoring of the model's performance during the training and ablation study phases.
- **Held-out**: This set is used only once to evaluate the final model after all training and ablations are complete.
This separation is designed to prevent inadvertent overfitting to the benchmarks during the iterative training process and to provide a more accurate measure of the model's generalization capabilities.
|**Category**|**Held-in**|**Held-out**|
|---|---|---|
|**General Knowledge**|KMMLU-Redux|KMMLU-HARD, KMMLU-Pro|
|**Reasoning**|MCLM|KSM, GPQA, AIME2024, AIME2025|
|**Korean-specific**|HAE-RAE Bench|CLIcK, KoBALT-700|
**Comparison with models trained on public datasets**
<table>
<thead>
<tr>
<th>Models</th>
<th># Instances</th>
<th>Methodology</th>
<th>Held-Out (Ko)</th>
<th>Held-Out (En)</th>
<th>Total</th>
</tr>
</thead>
<tbody>
<tr>
<th>KO-REASon-AX3_1-7B-0831(KONI-7B-R-20250831; Ours)</th>
<td>260k</td>
<td>SFT</td>
<td><b>44.6</b></td>
<td>41.2</td>
<td><u>43.3</u></td>
</tr>
<tr>
<th>KO-REASon-7B-Q2_5-0831(Ours)</th>
<td>260k</td>
<td>SFT</td>
<td><b>45.10</b></td>
<td>38.75</td>
<td><u>49.95</u></td>
</tr>
<tr>
<th>KO-REAson-KL3_1-8B-0831(KONI-Llama3.1-8B-R-20250831)</th>
<td>260k</td>
<td>SFT</td>
<td>40.13</td>
<td>30.57</td>
<td>43.66</td>
</tr>
<tr>
<td colspan="6" style="text-align:center; font-weight:bold;">Open Recipe (En)</td>
</tr>
<tr>
<th>OpenThinker3-7B</th>
<td>1.2M</td>
<td>SFT</td>
<td>33.6</td>
<td><b>55.5</b></td>
<td>41.8</td>
</tr>
<tr>
<th>s1.1-7B</th>
<td>1k</td>
<td>SFT</td>
<td>35.6</td>
<td>23.4</td>
<td>31.1</td>
</tr>
<tr>
<th>Llama-3.1-Nemotron-Nano-8B-v1</th>
<td>>3M</td>
<td>SFT & RL</td>
<td>27.0</td>
<td>44.1</td>
<td>33.4</td>
</tr>
<tr>
<td colspan="6" style="text-align:center; font-weight:bold;">Open Recipe (Ko)</td>
</tr>
<tr>
<th>Ko-R1-14B</th>
<td>45k</td>
<td>SFT</td>
<td><u>43.7</u></td>
<td><u>46.3</u></td>
<td><b>44.7</b></td>
</tr>
<tr>
<th>Ko-R1-7B</th>
<td>45k</td>
<td>SFT</td>
<td>27.3</td>
<td>36.1</td>
<td>30.6</td>
</tr>
<tr>
<th>LLaMa-3.1-Ko-Reasoning-8B</th>
<td>63k</td>
<td>SFT</td>
<td>17.7</td>
<td>7.7</td>
<td>14.0</td>
</tr>
</tbody>
</table>
**Held-out benchmark performance**
<table border="1" cellspacing="0" cellpadding="6">
<thead>
<tr>
<th rowspan="2">Model</th>
<th rowspan="2">Model Size</th>
<th colspan="2">General</th>
<th colspan="4">Reasoning</th>
<th colspan="2">Korean-Specific</th>
<th rowspan="2">Average<br>(Held-out)</th>
<th rowspan="2">Average<br>(Held-out-Ko)</th>
</tr>
<tr>
<th>KMMLU-HARD</th>
<th>KMMLU-Pro</th>
<th>KSM</th>
<th>AIME 2024</th>
<th>AIME 2025</th>
<th>GPQA</th>
<th>CLIcK</th>
<th>KoBALT-700</th>
</tr>
</thead>
<tbody>
<tr>
<td><b>Llama-3.1-Nemotron-Nano-8B</b></td>
<td>8.03</td><td>21.47</td><td>22.89</td><td>47.06</td><td>56.67</td><td>43.33</td><td>32.32</td><td>34.54</td><td>9.29</td><td>33.45</td><td>27.05</td>
</tr>
<tr>
<td><b>Llama-3.1-Korean-Reasoning-8B-Instruct</b></td>
<td>8.03</td><td>14.91</td><td>21.72</td><td>6.09</td><td>0.00</td><td>0.00</td><td>23.23</td><td>39.65</td><td>6.14</td><td>13.97</td><td>17.70</td>
</tr>
<tr>
<td><b>EXAONE-Deep-7.8B</b></td>
<td>7.82</td><td><u>40.96</u></td><td>37.35</td><td><b>70.80</b></td><td><b>70.00</b></td><td><b>63.33</b></td><td><b>64.65</b></td><td>54.24</td><td>18.86</td><td><b>52.52</b></td><td>44.44</td>
</tr>
<tr>
<td><b>DeepSeek-R1-Distill-Qwen-7B</b></td>
<td>7.62</td><td>0.00</td><td>23.00</td><td>56.09</td><td>60.00</td><td>40.00</td><td>43.43</td><td>0.00</td><td>8.29</td><td>28.85</td><td>17.48</td>
</tr>
<tr>
<td><b>DeepSeek-R1-Distill-Llama-8B</b></td>
<td>8.03</td><td>23.22</td><td>26.26</td><td>29.97</td><td>33.33</td><td>20.00</td><td><U>46.46</u></td><td>39.05</td><td>13.29</td><td>28.95</td><td>26.36</td>
</tr>
<tr>
<td><b>s1.1-7B</b></td>
<td>7.62</td><td>31.16</td><td><u>37.70</u></td><td>30.60</td><td>16.67</td><td>23.33</td><td>30.30</td><td><u>56.84</u></td><td><u>21.86</u></td><td>31.06</td><td>35.63</td>
</tr>
<tr>
<td><b>OpenThinker3-7B</b></td>
<td>7.62</td><td>30.31</td><td>26.26</td><td><u>63.59</u></td><td><u>66.67</u></td><td><u>53.33</u></td><td><u>46.46</u></td><td>47.69</td><td>10.14</td><td>35.63</td><td>30.60</td>
</tr>
<tr>
<td><b>Ko-R1-7B</b></td>
<td>7.61</td><td>28.46</td><td>19.31</td><td>51.61</td><td>46.67</td><td>33.33</td><td>28.28</td><td>32.48</td><td>4.71</td><td>30.61</td><td>27.31</td>
</tr>
<tr>
<td><b>KO-REAson-KL3_1-8B-0831(KONI-Llama3.1-8B-R-20250831)</b></td>
<td>8.03</td><td>44.64</td><td>40.08</td><td>37.96</td><td>23.33</td><td>30.00</td><td>38.38</td><td>56.39</td><td>21.57</td><td>30.57</td><td>40.13</td>
</tr>
<tr>
<td><b>KO-REASon-AX3_1-7B-0831 (KONI-7B-R-20250831)</b></td>
<td>7.26</td><td>45.57</td><td>38.13</td><td>52.80</td><td>53.33</td><td>33.33</td><td>36.87</td><td><b>62.86</b></td><td>23.43</td><td><u>43.29</u></td><td><u>44.56</u></td>
</tr>
<tr>
<td><b>KO-REASon-7B-Q2_5-0831</b></td>
<td>7.26</td><td><b>46.81</b></td><td><b>44.93</b></td><td>48.11</td><td>43.33</td><td>30.00</td><td>42.93</td><td>60.65</td><td><b>25.00</b></td><td>42.72</td><td><b>45.10</b></td>
</tr>
</tbody>
</table>
## Citation
```
The paper will be released soon!
```
## Contact
For any questions contact us via the following email :)
```
spthsrbwls123@yonsei.ac.kr
```
## Acknowlegments
This research was supported by the Korea Institute of Science and Technology Information (KISTI) (No.(KISTI) K25L1M1C1), aimed at developing KONI (KISTI Open Neural Intelligence), a large language model specialized in science and technology. |