ctc-finetune-recommendation.yaml

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

model:
  name: "omniASR_CTC_300M" # omniASR_CTC_1B, omniASR_CTC_3B

dataset:
  name: "example_dataset"
  train_split: "train"
  valid_split: "dev"
  storage_mode: "MIXTURE_PARQUET"
  task_mode: "ASR"
  mixture_parquet_storage_config:
    dataset_summary_path: "/path/to/your/dataset/language_distribution_0.tsv"
    beta_corpus: 0.5
    beta_language: 0.5
    fragment_loading:
      cache: True
  asr_task_config:
     max_audio_len: 960_000
     max_num_elements: 7_680_000
     batch_shuffle_window: 1
     normalize_audio: true
     example_shuffle_window: 0 # full-batch shuffling by default

tokenizer:
  name: "omniASR_tokenizer"

optimizer:
  config:
    lr: 1e-05

trainer:
  freeze_encoder_for_n_steps: 0
  mixed_precision:
    dtype: "torch.bfloat16"
  grad_accumulation:
    num_batches: 1 # tune this parameter if you are getting OOM

regime:
  num_steps: 5_000
  validate_every_n_steps: 500
  validate_after_n_steps: 500
  checkpoint_every_n_steps: 500
  checkpoint_after_n_steps: 500
  publish_metrics_every_n_steps: 500
  publish_metrics_after_n_steps: 500