# Quintus Distillation Pipeline
# Run profile: online full-vocabulary KD, 8B teacher -> 1.7B-Base student.
# Data: ~90K English-only samples from DistilQwen_100k.

data:
  dataset_path: "<REDACTED_ON_PURPOSE>"
  num_samples: 90234
  max_seq_len: 4096
  stream_shuffle_buffer_size: 20000
  stream_shuffle_seed: 25

model:
  teacher: "Qwen/Qwen3-8B"
  student: "Qwen/Qwen3-1.7B-Base"

  # The instruct tokenizer carries the chat template used to format the base
  # student into assistant-style training examples.
  tokenizer: "Qwen/Qwen3-1.7B"

  teacher_revision: "main"
  student_revision: "main"
  tokenizer_revision: "main"
  allow_remote_code: false

training:
  # Schedule
  num_epochs: 1
  validation_ratio: 0.02
  split_seed: 25

  # Optimizer
  learning_rate: 5.0e-6
  weight_decay: 0.1
  warmup_ratio: 0.05

  # Loss mix used by src/losses.py:
  # total = alpha * CE + (1 - alpha) * KD
  alpha: 0.3
  temperature: 2.0

  # Online KD streams full-vocabulary teacher logits. top_k is retained for
  # offline-KD compatibility/provenance checks.
  top_k: 8
  online_kd_token_chunk_size: 2048

  # Conservative B200 profile. Effective batch = 4 * 2 = 8.
  # If VRAM headroom is comfortable and Liger is installed, try 8 * 1.
  micro_batch_size: 4
  grad_accum_steps: 2
  gradient_checkpointing: false
  compile_model: false
  fused_adamw: true

  dataloader_workers: 8
  prefetch_factor: 2

  sequence_packing:
    enabled: true
    pack_length: 4096
    mask_first_token_after_separator: true

hub:
  # Prefer HF_TOKEN or huggingface-cli login for real runs.
  token: null
  username: "<REDACTED_ON_PURPOSE>"
  repo_name: "<REDACTED_ON_PURPOSE>"

paths:
  teacher_dir: "<REDACTED_ON_PURPOSE>"
  student_dir: "<REDACTED_ON_PURPOSE>"
  tokenizer_dir: "<REDACTED_ON_PURPOSE>"
  tokenized_dir: "<REDACTED_ON_PURPOSE>"
  logits_dir: "<REDACTED_ON_PURPOSE>"
  distilled_dir: "<REDACTED_ON_PURPOSE>"
  log_file: "<REDACTED_ON_PURPOSE>"
  system_info: "<REDACTED_ON_PURPOSE>"
  loss_csv: "<REDACTED_ON_PURPOSE>"