|
--- |
|
license: apache-2.0 |
|
language: |
|
- ko |
|
--- |
|
# KULLM-Polyglot-12.8B-v2 |
|
|
|
This model is a fine-tuned version of [EleutherAI/polyglot-ko-12.8b](https://huggingface.co/EleutherAI/polyglot-ko-12.8b) on a KULLM v2 |
|
|
|
Detail Codes are available at [KULLM Github Repository](https://github.com/nlpai-lab/KULLM) |
|
|
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
|
|
- learning_rate: 3e-4 |
|
- train_batch_size: 6 |
|
- seed: 42 |
|
- distributed_type: multi-GPU (A100 80G) |
|
- num_devices: 4 |
|
- gradient_accumulation_steps: 21.3 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- num_epochs: 8.0 |
|
|
|
### Framework versions |
|
|
|
- Transformers 4.28.1 |
|
- Pytorch 2.0.0+cu117 |
|
- Datasets 2.11.0 |
|
- Tokenizers 0.13.3 |