configs/phase2_8state_s020_v1.yaml

project:
  seed: 12345
  resolution: [512, 512]
  output_dir: artifacts/phase2_balanced_4000_8state

paths:
  comfyui: vendor/ComfyUI
  ideogram4: vendor/ideogram4
  stock_text_encoder: checkpoints/text_encoders/qwen3vl_8b_fp8_scaled.safetensors
  conditional: checkpoints/diffusion_models/ideogram4_fp8_scaled.safetensors
  teacher_pairs: artifacts/civitai_balanced_teacher_generation/nemotron_3_ultra_nvfp4_v1/teacher_pairs_4000.jsonl
  holdout_approval: data/civitai_balanced/teacher_v1/teacher_pairs_sfw_r_x_xxx_4000.approval.json
  training_index: artifacts/phase2_balanced_4000_training_dataset/index.jsonl

teacher:
  provider: fireworks
  base_url: https://api.fireworks.ai/inference/v1
  model: accounts/fireworks/models/nemotron-3-ultra-nvfp4
  temperature: 0.0
  max_tokens: 16384
  reasoning_effort: high
  aspect_ratio_placeholder: "1:1"
  strip_bboxes: true

conditioning:
  dtype: float16
  taps: [0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 35]
  cache_dir: artifacts/phase2_balanced_4000_teacher_conditioning

trajectory:
  # Full-trajectory contract: eight evenly distributed pre-step states from
  # V4_DEFAULT_20. The computed capture steps are [19, 16, 14, 11, 8, 5, 3, 0].
  # Cache manifests are checked before model load, so one-state data is rejected.
  state_mode: trajectory
  states_per_prompt: 8
  sampling_steps: 20
  sampler_preset: V4_DEFAULT_20
  seed: 12345
  cache_dir: artifacts/phase2_balanced_4000_teacher_trajectories_8state_debanner_s020_v1
  generation_batch_size: 1
  parallel_cfg: true
  correction:
    tensor: checkpoints/debanner/ideogram4_correction_v1.safetensors
    metadata: checkpoints/debanner/ideogram4_correction_v1.json
    # Same authoritative direction and operation, explicitly downscaled for
    # distillation after the 0.6 target produced an excessive first-state bias.
    strength: 0.2

training:
  # Adapter geometry is invariant across all production runs: rank/alpha 64,
  # all seven projection families, and all 36 Qwen blocks (inclusive range).
  lora_rank: 64
  lora_alpha: 64
  qwen_blocks: [0, 35]
  deployment_quantization: scaled_fp8_ste
  batch_size: 2
  gradient_accumulation: 8
  adaptive_behavior_batching:
    enabled: true
    max_padded_student_tokens: 1017
    first_oom_student_tokens: 1223
    benchmark_dir: artifacts/phase2_microbatch_cutoff
  optimizer: adamw
  optimizer_state_offload: cpu
  prune_unused_multimodal: true
  learning_rate: 0.00032
  weight_decay: 0.01
  default_session_steps: 50
  lr_control:
    mode: windowed_descent_rollback
    factor: 0.5
    minimum_learning_rate: 0.000001
    window_steps: 50
    probe_every_steps: 10
    confirmation_steps: 20
    minimum_descent_to_noise: 1.0
    minimum_relative_descent: 0.0
  validation_every: 50
  validation_states_per_record: 1
  gradient_clip_norm: 1.0
  lambda_velocity: 1.0
  lambda_states: 0.0

checkpointing:
  directory: artifacts/phase2_balanced_4000_8state_debanner_s020_v1_checkpoints
  every_optimizer_steps: 50
  save_first_optimizer_step: true
  save_on_interrupt: true
  resume: auto
  optimizer_boundary_only: true