wandb:
  entity: "vikhr-audio"
  project: "Borealis"

datasets:
  train:
    - name: "Vikhrmodels/ToneBooksPlus"
      split: "train"
      columns: ["audio", "text"]  
    - name: "Vikhrmodels/ToneSpeak"
      split: "train"
      columns: ["audio", "text"]
    - name: "Vikhrmodels/ReadyFormatDF2"
      split: "train"
      columns: ["audio", "text"]
    - name: "Vikhrmodels/ToneRuLS"
      split: "train"
      columns: ["audio", "text"]
    - name: "Vikhrmodels/ToneSlavic"
      split: "train"
      columns: ["audio", "sentence"] 
      rename_text: "sentence" 
      filter:
        locale_ru: true
    - name: "Vikhrmodels/ToneRuDevices"
      split: "train"
      columns: ["audio", "text"]
    - name: "Vikhrmodels/ReadyFormatDF"
      split: "train"
      columns: ["audio", "text"]
    - name: "Vikhrmodels/ToneRuDevicesAudiobooks"
      split: "train"
      columns: ["audio", "text"]
    - name: "bond005/podlodka_speech"
      split: "train"
      columns: ["audio", "transcription"] 
      rename_text: "transcription"  
  #    - name: "Vikhrmodels/ToneGolosOpus"
   #   config: "Crowd"
    #  split: "train"
    #  columns: ["audio", "text"]
    - name: "Vikhrmodels/ToneGolosOpus"
      config: "Farfield"
      split: "train"
      columns: ["audio", "text"]
  
  val:
    - name: "Vikhrmodels/ToneBooksPlus"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ToneSpeak"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ReadyFormatDF2"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ToneRuLS"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ToneSlavic"
      split: "validation"
      columns: ["audio", "sentence"]
      rename_text: "sentence"
      select_range: 279
      filter:
        locale_ru: true
    - name: "Vikhrmodels/ToneRuDevices"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ReadyFormatDF"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ToneRuDevicesAudiobooks"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "bond005/podlodka_speech"
      split: "validation"
      columns: ["audio", "transcription"]
      rename_text: "transcription"
      select_range: 20
    - name: "Vikhrmodels/ToneGolosOpus"
      config: "Crowd"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279
    - name: "Vikhrmodels/ToneGolosOpus"
      config: "Farfield"
      split: "validation"
      columns: ["audio", "text"]
      select_range: 279

  noise:
    name: "Vikhrmodels/Audio_Noise_Dataset"
    split: "Musan"
  ir:
    name: "Vikhrmodels/Audio_Noise_Dataset"
    split: "Echo"

  num_proc: 8
  sampling_rate: 16000
  clean_dataset: true

model:
  whisper:
    pretrained: "openai/whisper-large-v3"
    dtype: "bfloat16"
    attn_implementation: "eager"
  language_model:
    pretrained: "Unsloth/Qwen3-4B"
    full_finetuning: true
    attn_implementation: "sdpa"
  special_tokens:
    start_audio: "<|start_of_audio|>"
    end_audio: "<|end_of_audio|>"
  max_text_len: 512

augmentation:
  sample_rate: 16000
  stages: "default" 

training:
  output_dir: "./asr_qwen_ckpts"
  per_device_train_batch_size: 16
  per_device_eval_batch_size: 16
  dataloader_num_workers: 64
  save_total_limit: 50
  num_train_epochs: 5
  warmup_ratio: 0.05
  learning_rate: 3e-4
  bf16: true
  eval_strategy: "steps"
  save_strategy: "steps"
  eval_steps: 15000000
  save_steps: 10000
  logging_steps: 50
  report_to: "wandb"
  save_safetensors: false
  optim: "adamw_torch_fused"
  ddp_find_unused_parameters: false
  lr_scheduler_type: "cosine"
  gradient_checkpointing_kwargs: {"use_reentrant":False}
  # Hugging Face Hub настройки
  push_to_hub: false
  hub_model_id: "Vikhrmodels/Borealis-ASR-4B"
  hub_strategy: "checkpoint"
  hub_private_repo: false
  hub_always_push: false
generation:
  max_new_tokens: 512
  do_sample: false
  #num_beams: 5
  early_stopping: true
  repetition_penalty: 1.2
  temperature: 0.79

metrics:
  random_samples: 5