ESPnet3 Publish Configuration

This page explains the publish.yaml schema used by pack/upload stages. It controls how model artifacts are assembled and optionally uploaded.

Minimum required keys (typical pack/upload)

Required for packing:

• pack_model.out_dir (output directory)
• pack_model.strategy (or rely on default if set elsewhere)

Required for upload:

• upload_model.hf_repo

Common optional:

• pack_model.infer_dir (to pull metrics into README)
• pack_model.include / pack_model.exclude
• pack_model.files / pack_model.yaml_files

Minimal pack example:

pack_model:
  strategy: auto
  out_dir: exp/model_pack

Minimal upload example:

upload_model:
  hf_repo: yourname/your-model-repo

✅ Config sections overview

Section	Description
pack_model	Packaging strategy and file selection.
upload_model	Upload destination for model artifacts.

Core config layout

pack_model:
  strategy: auto  # auto|espnet2|espnet3

  # espnet2 packer only
  task: asr
  out_dir: exp/model_pack

  # used to locate scores.json from the measure stage
  # (note: some pipelines may write measures.json instead)
  infer_dir: ${exp_dir}/infer

  # extra files/dirs outside exp/<exp_tag> to include in the archive
  include:
    - ${load_yaml:conf/train.yaml,tokenizer.save_path}
  exclude:
    - "**/*.log"
    - "**/tensorboard/**"
    - "**/wandb/**"

  # used to generate meta.yaml (for espnet2-style from_pretrained)
  files:
    asr_model_file: ${exp_dir}/valid.acc.best.pth
    lm_file: ${exp_dir}/lm.pth
    ngram_file: ${exp_dir}/ngram.arpa
    bpemodel: ${exp_dir}/tokenizer/bpe.model

  # used to generate meta.yaml (for espnet2-style from_pretrained)
  yaml_files:
    asr_train_config: ${exp_dir}/config.yaml
    lm_train_config: ${exp_dir}/lm_config.yaml

upload_model:
  hf_repo: yourname/your-model-repo

Notes

• upload_model uses huggingface-cli, so login beforehand (huggingface-cli login) or set an API token via HF_TOKEN in the environment.