This model is a instruct-tuned ko-ref-llama2-7b model, using only 10% of [Kullm, OIG, KoAlpaca] Instruction dataset.
len10_k100_mppl_n0.1.json -> 152step
Training hyperparameters
learning_rate: 5e-5
train_batch_size: 1
seed: 42
distributed_type: multi-GPU (A30 24G) + CPU Offloading(160GB)
num_devices: 2
gradient_accumulation_steps: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
