taeminlee's picture
Create README.md
ffe1905
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