run_name: AMD-OLMo-1B-SFT-1st-phase
seed: 6198
dry_run: false

wandb:
  name: ${run_name}
  project: AMD-OLMo
  group: SFT

model:
  d_model: 2048
  n_heads: 16
  n_layers: 16
  mlp_ratio: 8
  weight_tying: true
  alibi: false
  rope: true
  flash_attention: false
  attention_dropout: 0.0
  attention_layer_norm: false
  multi_query_attention: false
  include_bias: false
  block_type: sequential
  layer_norm_type: default
  layer_norm_with_affine: false
  bias_for_layer_norm: false
  attention_layer_norm_with_affine: false
  activation_type: swiglu
  residual_dropout: 0.0
  embedding_dropout: 0.0
  max_sequence_length: 2048
  vocab_size: 50280
  embedding_size: 50304
  eos_token_id: 50279
  pad_token_id: 1
  init_device: meta
  init_fn: mitchell

compile:
  fullgraph: false

optimizer:
  name: adamw
  learning_rate: 2.0e-5
  weight_decay: 0
  betas:
  - 0.9
  - 0.95
  metrics_log_interval: 10

scheduler:
  name: linear_with_warmup
  t_warmup: 200
  alpha_f: 0.001

tokenizer:
  identifier: tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json
  truncate_direction: right

save_folder: ./outputs/${run_name}/
save_overwrite: true
# Sharded checkpoints (best for restarts)
save_interval: 1000
save_num_checkpoints_to_keep: -1
# Unsharded checkpoints (for final storage)
save_interval_unsharded: 10000
save_num_unsharded_checkpoints_to_keep: -1

load_path: path_to_unsharded_pretrain_checkpoint
reset_trainer_state: true

max_duration: 3ep  # train 3 epochs
global_train_batch_size: 128
device_train_microbatch_size: 8

precision: amp_bf16

fsdp:
  wrapping_strategy: null
  precision: mixed

max_grad_norm: 1.0
max_grad_norm_ratio: null

speed_monitor:
  window_size: 20

eval_interval: ${save_interval}
eval_subset_num_batches: -1
device_eval_batch_size: ${device_train_microbatch_size}
evaluators:
  - label: piqa
    type: downstream

  - label: hellaswag
    type: downstream

  - label: winogrande
    type: downstream

  - label: openbook_qa
    type: downstream

  # - label: boolq  # requires implemention of the pmi_dc matrix
  # type: downstream

  - label: sciq
    type: downstream

  - label: arc_easy
    type: downstream

  # - label: arc_challenge  # requires implemention of the pmi_dc matrix
  #   type: downstream

  - label: copa
    type: downstream

  - label: rte
    type: downstream

  - label: commitment_bank
    type: downstream

  - label: mrpc
    type: downstream

  - label: sst2
    type: downstream

data:
  pad_direction: right
  num_workers: 0
  drop_last: true
  pin_memory: true
  prefetch_factor: 1
  persistent_workers: true
  timeout: 0
  generate_attention_mask: true
  paths:
    - ./datasets/tulu/input_ids.npy
  label_mask_paths:
    - ./datasets/tulu/label_mask.npy