metadata
license: cc-by-4.0
KoQuality-Polyglot-5.8b
KoQuality-Polyglot-5.8b is an auto-regressive language model that conducts instruction tuning with KoQuality datasets on Polyglot-5.8b model. Our best model is trained on KoQuality dataset, which is curated by proposed method (len_group=5, k=100, method=ppl_sampling).
Average accuracy score of the KoBEST datasets
We use KoBEST benchmark datasets(KoBEST_boolq, KoBEST_copa, KoBEST_hellaswag, KoBEST_sentineg, KoBEST_wic) to compare the performance of our best model and other models accuracy. Our model outperforms other models in the average accuracy score of the KoBEST datasets.
| Model | 0-shot | 1-shot | 2-shot | 5-shot | 10-shot |
|---|---|---|---|---|---|
| koquality-polyglot-5.8b | 0.5472 | 0.5979 | 0.6260 | 0.6486 | 0.6535 |
| polyglot-ko-5.8b | 0.5587 | 0.5977 | 0.6138 | 0.6431 | 0.6457 |
| koalpcaca-polyglot-5.8b | 0.5085 | 0.5561 | 0.5768 | 0.6097 | 0.6059 |
| kullm-polyglot-5.8b | 0.5409 | 0.6072 | 0.5945 | 0.6345 | 0.6530 |
Training hyperparameters
- learning_rate: 5e-5
- train_batch_size: 4
- seed: 42
- distributed_type: multi-GPU (A100 80G)
- num_devices: 4
- gradient_accumulation_steps: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 2.0
Citation
@misc{2023koqaulity,
title = {KoQuality: Curation of High-quality Instruction Data for Korean Language Models},
author = {Na, Yohan and Kim, Dahye and Chae, Dong-Kyu},
journal={Proceedings of the 35th Annual Conference on Human and Cognitive Language Technology (HCLT 2023)},
pages={},
year = {2023},
}