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>&gt;3M</td>
      <td>SFT &amp; 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.