data:
  augment:
    elastic: 0.3
    illumination_variation: true
    rotation: 5
  tlhdig_pairs: datasets/processed/tlhdig_photo_pairs.jsonl
  tokenizer:
    add_tokens:
    - <SIGN>
    - </SIGN>
    - <LINE>
    - </LINE>
    - <BROKEN>
    - <UNCERTAIN>
    base: ByT5ByteTokenizer
    include_phonetic: true
  train_manifest: datasets/sources/hitit_local/manifest_detection.parquet
encoder_decoder:
  cross_attention_heads: 8
  cross_attention_layers: 4
generation:
  beam_size: 5
  early_stopping: true
  no_repeat_ngram_size: 3
lm_fusion:
  enabled: true
  kenlm_binary: hitit_ocr/runs/hitit_kenlm_5gram.binary
  lambda_lm: 0.3
loss:
  auxiliary:
    confidence_penalty: 0.1
    ctc_alignment: 0.3
  primary: cross_entropy
model:
  architecture: vision_encoder_decoder
  decoder:
    length_penalty: 0.6
    max_length: 512
    num_beams: 5
    pretrained: google/byt5-small
    type: byt5-small
  encoder:
    image_size: 384
    lora_alpha: 64
    lora_r: 32
    patch_size: 14
    pretrained: hitit_ocr/runs/ssl_dinov3_continual/checkpoint.pt
    type: dinov3_vitl14
output_dir: hitit_ocr/runs/seq2seq_hitit_v1/
targets:
  CER: 0.05
  chrF: 0.95
  sign_level_accuracy_derived: 0.94
training:
  batch_size: 8
  epochs: 3
  fp16: true
  gradient_accumulation: 4
  gradient_checkpointing: true
  label_smoothing: 0.1
  lr_decoder: 0.0001
  lr_encoder: 5.0e-06
  teacher_forcing_prob: 0.9
  warmup_steps: 2000