# test_pli_wholeword.yml — Phase P2: PLI whole-word evaluation

project_name: "deep-stylometry"

model:
  base_checkpoint: "answerdotai/ModernBERT-base"
  pooling_method: "pli"
  skip_list: true
  lm_hidden_size: 768
  patch_method: "wholeword"
  patch_size: 1
  patch_compression: "mean"
  patch_cross_attn_dim: 128
  patch_cross_attn_heads: 2

data:
  ds_name: "halvest"
  batch_size: 64
  tokenizer_name: "answerdotai/ModernBERT-base"
  max_length: 512
  padding: "do_not_pad"
  truncation: "longest_first"
  add_special_tokens: true
  map_batch_size: 1000
  load_from_cache_file: true
  subsets: ["base-2", "base-4", "base-6", "base-8", "base-10"]
  shuffle: true
  test_subset: "base-10"

test:
  device: "gpu"
  num_devices: 1
  log_every_n_steps: 1
  precision: "16-mixed"
  use_wandb: true