# Model defaults for unsloth/gte-modernbert-base
# Based on ModernBert.py embedding notebook

training:
  max_seq_length: 512
  # num_epochs: 2
  num_epochs: 0
  learning_rate: 3e-5
  batch_size: 256
  gradient_accumulation_steps: 1
  warmup_ratio: 0.03
  max_steps: 30
  save_steps: 30
  weight_decay: 0.001
  random_seed: 3407
  packing: false
  train_on_completions: false
  gradient_checkpointing: "unsloth"
  optim: "adamw_8bit"
  lr_scheduler_type: "constant_with_warmup"

lora:
  lora_r: 64
  lora_alpha: 128
  lora_dropout: 0.0
  target_modules:
    - "Wi"
    - "Wo"
    - "Wqkv"
  use_rslora: false
  use_loftq: false

logging:
  enable_wandb: false
  wandb_project: "embedding-finetuning"
  enable_tensorboard: false
  tensorboard_dir: "runs"
  log_frequency: 50