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