code/configs/classification_hitit_only.yaml

# HITIT-ONLY Classification (v3) — Target top-1 ≥ 90%
# 198 sınıf, 21K kayıt, IR=1229×, manuel etiketli
# NOT: Hierarchical head gereksiz (198 flat yeter)

backbone:
  arch: dinov3_vitl14        # LARGE — 300M params
  pretrained: hitit_ocr/runs/ssl_dinov3_continual/checkpoint.pt  # SSL continual
  freeze: false

lora:
  r: 32                       # Larger rank (198 class için yeter)
  alpha: 64
  dropout: 0.1
  targets: ['qkv', 'proj', 'mlp.fc1', 'mlp.fc2']
  bias: none

head:
  type: flat                  # NO hierarchy
  n_classes: 198              # Hitit direct
  dropout: 0.3

loss:
  type: cross_entropy
  label_smoothing: 0.1
  class_weights: sqrt_inverse
  class_weight_cap: 10.0      # Avoid extreme rare tier boost

sampler:
  type: weighted_random
  weight_mode: sqrt_inverse
  replacement: true

augmentation:
  mixup: {alpha: 0.2, p: 0.5}
  cutmix: {alpha: 1.0, p: 0.3}
  elastic_transform: {alpha: 40.0, sigma: 6.0, p: 0.5}
  grid_distortion: {num_steps: 5, distort_limit: 0.15, p: 0.3}
  rotation: {degrees: 5, p: 0.5}
  color_jitter: {brightness: 0.2, contrast: 0.2, p: 0.5}
  horizontal_flip: {enabled: false}        # Cuneiform yön-duyarlı!
  illumination_variation: {enabled: true, p: 0.3}   # MaiCuBeDa stili

optimizer:
  type: AdamW
  lrs:
    backbone: 1.0e-5
    lora: 5.0e-4
    head: 1.0e-3
  weight_decay: 0.05
  betas: [0.9, 0.999]

scheduler:
  type: cosine_with_warmup
  warmup_epochs: 5
  total_epochs: 100
  eta_min: 1.0e-6

ema:
  enabled: true
  decay: 0.9999
  start_epoch: 20

swa:
  enabled: true
  start_epoch: 85           # son 15 epoch
  swa_lr: 5.0e-5
  update_bn_on_end: true

batch_size: 64
gradient_accumulation: 2    # effective 128
epochs: 100                 # LP 10 + FT 90 with strong aug
gradient_clip_val: 1.0

# Mixed precision (Option A — bf16 daha stabil, aynı hız)
bf16: true                  # bit-exact accuracy, fp16'dan stabil
fp16: false

# Training efficiency (Option A — accuracy-neutral speedups)
torch_compile:
  enabled: true
  mode: "max-autotune"      # 1.3x throughput, bit-exact
flash_attention:
  enabled: true
  version: 2                # FA2 on A100
fused_optimizer:
  enabled: true             # AdamW(fused=True), 1.1x

normalize:
  mean: [0.489, 0.448, 0.424]   # dataset-specific
  std: [0.362, 0.359, 0.364]

data:
  manifest: datasets/sources/hitit_local/manifest_classification.parquet
  filter: "class_sample_count >= 5 AND integrity_ok == True AND quality_gate_pass == True"
  # 177 sınıf, 21162 sample (99.8% coverage) — 21 rare sınıf hariç
  cv:
    strategy: tablet_view_fold
    val_fold: 0
    test_fold: 4               # LOCKED for final eval
  # Rendered domain adaptation: MaiCuBeDa Hitit-uyumlu augment eklenebilir
  aux_domain_adapt:
    enabled: true
    sources: ['maicubeda', 'heicubeda']  # rendered → hitit style
    weight: 0.1                           # %10 karışım

eval:
  metrics: [top1_acc, top5_acc, macro_f1, balanced_acc]
  stratify_by: class_frequency_tier
  per_tablet_id: true                     # tablet-level aggregation
  selective_classification:
    enabled: true
    thresholds: [0.3, 0.5, 0.6, 0.7, 0.8]

targets:
  top1_accuracy: 0.90          # ⭐ PRIMARY TARGET
  top5_accuracy: 0.98
  macro_f1: 0.88
  selective_acc_90cov: 0.94

output_dir: hitit_ocr/runs/hitit_only_dinov3/