KO-REAson-G3-4B-0831

KO-REAson

KO-REAson is a series of Korean-centric reasoning language models developed in collaboration with OneLineAI, KISTI-KONI, HAE-RAE 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) and 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.

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)

• 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.

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 :)

[email protected]

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.