shared:
  project_name: continual_prompt_pretrained_mlp
  use_tqdm: true
  random_seed: 42

default: &default
  model_name: google/t5-large-lm-adapt
  wandb_name: null
  train_batch_size: 32
  valid_batch_size: 32
  num_epochs: 100
  peft_params: null  # no mutation
  hot_modules: null  # fine-tune all
  balancify_train: false
  best_finder:
    save: true
    metric: valid_f1-score-ma
    higher_better: true
  tasks:
  - glue:cola

run_configs:
# - <<: *default
#   wandb_name: large_5t_mlp128
#   learning_rate: 0.02
#   hot_modules:
#   - sadcl_learned_embeddin
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 5
#     mlp_emb: 128
    
# - <<: *default
#   wandb_name: large_10t_mlp128
#   learning_rate: 0.02
#   hot_modules:
#   - sadcl_learned_embeddin
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 10
#     mlp_emb: 128
    
# - <<: *default
#   wandb_name: large_5t_mlp128_not_freeze
#   learning_rate: 0.02
#   hot_modules:
#   - sadcl
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 5
#     mlp_emb: 128
    
# - <<: *default
#   wandb_name: large_10t_mlp128_not_freeze
#   learning_rate: 0.02
#   hot_modules:
#   - sadcl
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 10
#     mlp_emb: 128
    
# - <<: *default
#   wandb_name: large_5t_mlp128_not_freeze_lowlr
#   learning_rate: 0.001
#   hot_modules:
#   - sadcl
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 5
#     mlp_emb: 128
    
# - <<: *default
#   wandb_name: large_10t_mlp128_not_freeze_lowlr
#   learning_rate: 0.001
#   hot_modules:
#   - sadcl
#   train_batch_size: 24
#   valid_batch_size: 24
#   peft_params:
#     kind: encoder_emb
#     n_tokens: 10
#     mlp_emb: 128

- <<: *default
  wandb_name: large_100t_mlp128_lr.02
  learning_rate: 0.02
  hot_modules:
  - sadcl_learned_embeddin
  train_batch_size: 24
  valid_batch_size: 24
  peft_params:
    kind: encoder_emb
    n_tokens: 100
    mlp_emb: 128