metadata
license: apache-2.0
datasets:
- nlpai-lab/kullm-v2
language:
- ko
KULLM-Polyglot-5.8B-v2
This model is a parameter-efficient fine-tuned version of EleutherAI/polyglot-ko-5.8b on a KULLM v2
Detail Codes are available at KULLM Github Repository
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3e-4
- train_batch_size: 128
- seed: 42
- distributed_type: multi-GPU (A100 80G)
- num_devices: 4
- gradient_accumulation_steps: 8
- 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