Tesser-Llama-3-Ko-8B

Model Details

This model was continually pretrained from the Meta-Llama-3-8B, using English and Korean datasets. The goal is to enhance its proficiency in Korean while maintaining its English language capabilities from the original model.

Datasets

We sampled 16B tokens from the following datasets for training:

<table> <tr> <td><strong>Sources</strong> </td> <td><strong>Tokens (Llama-3-8B)</strong> </td> </tr> <tr> <td>AI-Hub </td> <td>9.2B </td> </tr> <tr> <td>Modu Corpus </td> <td>5.8B </td> </tr> <tr> <td>Wikipedia </td> <td>5.4B </td> </tr> </table>

Hyperparameters

<table> <tr> <td><strong>Learning rate</strong></td> <td><strong>Optimizer</strong></td> <td><strong>Betas</strong></td> <td><strong>Weight decay</strong></td> <td><strong>Warm-up ratio</strong></td> </tr> <tr> <td>3e-5</td> <td>AdamW</td> <td>(0.9, 0.95)</td> <td>0.1</td> <td>0.05</td> </tr> </table>

Intended Use

This model has not been fine-tuned, so you will need to train it on your own dataset before using it.

Evaluations

We evaluated this model using both English and Korean benchmarks, and compared it with similar models that were continually pretrained from the Meta-Llama-3-8B.

<table> <tr> <td></td> <td colspan="4"><strong>English</strong></td> <td colspan="3"><strong>Korean</strong></td> </tr> <tr> <td><strong>Model</strong></td> <td><strong>MMLU (5-shot)</strong></td> <td><strong>HellaSwag (10-shot)</strong></td> <td><strong>GSM8K (8-shot, CoT)</strong></td> <td><strong>BBH (3-shot, CoT)</strong></td> <td><strong>KMMLU (5-shot)</strong></td> <td><strong>HAE-RAE (5-shot)</strong></td> <td><strong>KoBEST (5-shot)</strong></td> </tr> <tr> <td>meta-llama/Meta-Llama-3-8B</td> <td><strong>65.1</strong></td> <td><strong>82.1</strong></td> <td><strong>52.0</strong></td> <td><strong>61.9</strong></td> <td>40.2</td> <td>61.1</td> <td>69.2</td> </tr> <tr> <td>saltlux/Ko-Llama3-Luxia-8B</td> <td>57.1</td> <td>77.1</td> <td>32.3</td> <td>51.8</td> <td>39.4</td> <td>69.2</td> <td>71.9</td> </tr> <tr> <td>beomi/Llama-3-Open-Ko-8B</td> <td>56.2</td> <td>77.4</td> <td>31.5</td> <td>46.8</td> <td>40.3</td> <td>68.1</td> <td><u>72.1</u></td> </tr> <tr> <td>beomi/Llama-3-KoEn-8B</td> <td>52.5</td> <td>77.7</td> <td>21.2</td> <td>43.2</td> <td><u>40.8</u></td> <td><u>71.3</u></td> <td><strong>73.8</strong></td> </tr> <tr> <td><strong>tesser-ai/Tesser-Llama-3-Ko-8B</strong></td> <td><u>60.5</u></td> <td><u>79.8</u></td> <td><u>40.3</u></td> <td><u>56.3</u></td> <td><strong>42.5</strong></td> <td><strong>72.1</strong></td> <td><strong>73.8</strong></td> </tr> </table>

Limitations

We trained this model using a context length of 4k due to resource limitations and to maximize training speed. However, the original model was trained with a context length of 8k, so an 8k context length could work well in downstream tasks.

License

This model follows the original Llama-3 license.