model_config: configs/model_500m.yaml
tokenizer_dir: artifacts/kawk500m-sft-dataset/tokenizer

data:
  sequence_length: 2048
  hf_dataset:
    repo_id: Infinity08/KAWK500M-Korean-SFT-v1
    config_name: default
    revision: 0a8e914359a063b916cfd9a3ee068a38ddcc1f79
    train_split: train
    validation_split: validation
    shuffle_buffer: 10000
    validation_sequences: 300

training:
  output_dir: runs/kawk500m-ko-instruct-v1-a100
  seed: 5150
  precision: bf16
  per_device_batch_size: 12
  gradient_accumulation_steps: 5
  epochs: 2.0
  max_steps: null
  learning_rate: 2.0e-5
  min_learning_rate: 2.0e-6
  warmup_ratio: 0.03
  weight_decay: 0.01
  adam_beta1: 0.9
  adam_beta2: 0.95
  adam_eps: 1.0e-8
  max_grad_norm: 1.0
  gradient_checkpointing: false
  log_every: 10
  eval_every: 100
  eval_batches: 20
  save_every: 250
  save_final: true
  torch_compile: false
  compile_mode: default