# Autoregressive (GPT-2 style) baseline config.
#
# Matches new_owt.yaml in scale / data / optimizer so results are directly
# comparable to the diffusion runs. Architecture is a standard decoder-only
# Transformer with RoPE and causal self-attention; no DiT / no adaLN.

model:
  vocab_size: 50257
  hidden_size: 768
  n_blocks: 12
  n_heads: 12
  max_seq_len: 512
  dropout: 0.0

training:
  seed: 0
  batch_size: 64
  num_steps: 500000
  lr: 5.0e-4
  lr_min: 5.0e-5
  warmup_steps: 10000
  weight_decay: 0.02
  grad_clip: 1.0
  adam_betas: [0.9, 0.99]
  adam_eps: 1.0e-9
  dtype: bf16
  compile: default          # set to "off" to disable torch.compile

  log_interval: 100
  eval_interval: 5000
  save_interval: 10000

data:
  dataset: openwebtext
  seq_len: 512
  cache_dir: data/owt_cache
  num_workers: 4
  max_train_samples: null
  max_val_samples: 100000
  mode: subsample           # 和 new_owt.yaml 对齐：1 doc/sample, BOS/EOS, random window + pad

logging:
  use_wandb: true
  project: sad_ar_baseline
  save_dir: outputs/ar_baseline