ESPnet3 Publication Configuration

This page explains the publication.yaml schema used by pack_model and upload_model.

Important

Publication config controls packaging and upload. Keep training and inference logic out of this file.

Minimum required keys

For packing:

• pack_model.out_dir

For upload:

• upload_model.hf_repo

Common optional keys:

• pack_model.inference_dir
• pack_model.readme
• pack_model.include
• pack_model.exclude
• pack_model.files

Basic shape

recipe_dir: .
exp_tag: ${self_name:}
exp_dir: exp/${exp_tag}
inference_dir: ${exp_dir}/inference

pack_model:
  out_dir: ${exp_dir}/model_pack
  inference_dir: ${inference_dir}
  readme: egs3/TEMPLATE/asr/src/hf_model_repo_readme_template.md
  include:
    - ${recipe_dir}/src
  exclude:
    - inference
    - inference_*
    - last.ckpt
  files:
    asr_model_file: ${exp_dir}/last.ckpt

upload_model:
  hf_repo: yourname/your-model-repo

Main sections

Section	Description
recipe_dir, exp_dir, inference_dir	Path scaffold used during packing
pack_model	Bundle creation settings
upload_model	Hugging Face upload settings

pack_model

pack_model first copies ${exp_dir}, then applies:

• include
• exclude
• config rewriting
• meta.yaml generation

Important fields:

• out_dir: where the packed model is written
• inference_dir: where metrics.json is read from for README/results
• readme: README template path
• include: extra files or directories to copy
• exclude: glob patterns to remove
• files: named artifacts written into meta.yaml

upload_model

upload_model.hf_repo is the main required field for upload.

The upload step uses the packed directory from pack_model.out_dir.

Note

pack_model may also receive training_config, inference_config, and metrics_config from run.py when those were loaded in the same run.

Naming note

Use inference_dir, not decode_dir. Current ESPnet3 publication config no longer uses decode naming for this path.

Related pages

Metrics Config

See how `metrics.json` is prepared before packaging.

Publish Stage

Read the stage-level flow for packing and uploading models.

Config Overview

Return to the full config map across the pipeline.