StyleInstructCapsDB / hf_release_for_spkcap_6e6bb298.plan.md
StyleInstructCaps's picture
Add files using upload-large-folder tool
563ce83 verified
|
Raw History Blame Contribute Delete
7.33 kB
metadata
name: HF release for spkcap
overview: >-
  Package the 7 source JSON manifests as a single Hugging Face dataset repo
  (CapSpeech-style) with 6 configs and parquet files under
  `/home/schand30/scratchbsisman/schand30/spkcap/release/`, drop local absolute
  `audio_path`s, and provide scripts to prepare and push the release without
  modifying any source files.
todos:
  - id: scaffold
    content: >-
      Create release/ skeleton: scripts/, data/, requirements.txt, README.md
      placeholder
    status: pending
  - id: prepare_script
    content: >-
      Write release/scripts/prepare_release.py that converts the 7 JSONs into
      parquet shards under release/data/<config>/, dropping local audio_path
    status: pending
  - id: readme
    content: >-
      Write release/README.md with the configs YAML front-matter and
      CapSpeech-style body (overview, usage, per-config schemas, audio source
      mapping)
    status: pending
  - id: upload_script
    content: >-
      Write release/scripts/upload_to_hf.py wrapper around
      huggingface_hub.upload_large_folder, parameterized by org/name
    status: pending
  - id: user_run
    content: >-
      User installs deps, runs prepare_release.py, fills in repo name, runs
      upload_to_hf.py
    status: pending
isProject: false

Goal

Release the seven manifests as a HF dataset repo modeled on OpenSound/CapSpeech: one repo, multiple configs, parquet shards, metadata-only (no audio), with a placeholder repo name (<HF_ORG>/<DATASET_NAME>) the user fills in before pushing.

All work happens under /home/schand30/scratchbsisman/schand30/spkcap/release/. Source files in data/data/ and data/evaluation/data_input_dir/ are read-only.

Config / split layout

Each source JSON has an {"annotation": [...]} wrapper with heterogeneous per-file schemas, so we use one HF "config" per logical dataset (this is what enables HF's split selector UI like CapSpeech's).

  • StyleInstructCaps
    • train <- data/data/train_fullscale_with_desc_v2.json (1,031,840 rows, 3.3 GB)
    • eval <- data/data/evaluation_holdout_with_desc_llama.json (14,756 rows)
  • OODPrompts
    • test <- data/evaluation/data_input_dir/vctk_manifest_2utt_36prompts_unseen.json (32 rows)
  • Eval500
    • test <- data/evaluation/data_input_dir/tbd1_500_llama_global.json (434 rows)
  • IDBalanced
    • test <- data/evaluation/data_input_dir/balance_acccent_gender_eval_set2_with_fields.json (400 rows)
  • UDVCTK
    • test <- data/evaluation/data_input_dir/vctk_manifest_10prompts.json (320 rows)
  • UnseenEval
    • test <- data/evaluation/data_input_dir/giga_manifest.json (600 rows)

Each row keeps all original fields except the local absolute audio_path (dropped). relative_audio_path is preserved as the portable identifier; the README documents which upstream dataset each source value maps to.

Output layout under release/

release/
├── README.md                       # dataset card with YAML front-matter declaring configs
├── scripts/
│   ├── requirements.txt            # huggingface_hub, datasets, pyarrow, tqdm
│   ├── prepare_release.py          # JSON -> parquet under release/data/<config>/
│   └── upload_to_hf.py             # pushes release/ to <HF_ORG>/<DATASET_NAME>
└── data/
    ├── StyleInstructCaps/
    │   ├── train-00000-of-000NN.parquet   # ~100k rows per shard => ~11 shards
    │   └── eval-00000-of-00001.parquet
    ├── OODPrompts/test-00000-of-00001.parquet
    ├── Eval500/test-00000-of-00001.parquet
    ├── IDBalanced/test-00000-of-00001.parquet
    ├── UDVCTK/test-00000-of-00001.parquet
    └── UnseenEval/test-00000-of-00001.parquet

README.md front-matter (mirrors CapSpeech "data files via YAML" pattern)

---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-to-speech
configs:
- config_name: StyleInstructCaps
  data_files:
  - split: train
    path: data/StyleInstructCaps/train-*.parquet
  - split: eval
    path: data/StyleInstructCaps/eval-*.parquet
- config_name: OODPrompts
  data_files:
  - split: test
    path: data/OODPrompts/test-*.parquet
- config_name: Eval500
  data_files:
  - split: test
    path: data/Eval500/test-*.parquet
- config_name: IDBalanced
  data_files:
  - split: test
    path: data/IDBalanced/test-*.parquet
- config_name: UDVCTK
  data_files:
  - split: test
    path: data/UDVCTK/test-*.parquet
- config_name: UnseenEval
  data_files:
  - split: test
    path: data/UnseenEval/test-*.parquet
---

The README body mirrors CapSpeech's structure: Overview, License (CC BY-NC 4.0 placeholder), Usage snippet (load_dataset("<HF_ORG>/<DATASET_NAME>", "StyleInstructCaps", split="train")), per-config Schema tables, Audio Sources table mapping each source value (emilia, voxceleb, VCTK, ears, eval_giga) to its upstream dataset URL, and a Citation block with <TODO citation>.

prepare_release.py behavior

Single, idempotent script that:

  1. Loads each source JSON via json.load(...)["annotation"]. Train file (3.3 GB) needs ~10-12 GB RAM at peak; the script logs a warning and recommends running on a node with >=16 GB free RAM. (If that's a concern we can swap to ijson streaming, but starts simpler.)
  2. Drops the audio_path key from every record (keeps relative_audio_path).
  3. Builds a datasets.Dataset per file (preserves nested lists/nullable fields like situational_tags).
  4. Writes parquet to the layout above. For StyleInstructCaps/train, shards via Dataset.shard(num_shards=ceil(N/100_000), index=i).to_parquet(...) so each shard is roughly 100k rows / a few hundred MB. All other splits write a single *-00000-of-00001.parquet.
  5. Prints final row counts to verify against the source counts (1031840 / 14756 / 32 / 434 / 400 / 320 / 600).

upload_to_hf.py behavior

Thin wrapper around huggingface_hub.HfApi:

  • Reads HF_ORG and DATASET_NAME from CLI args (or env).
  • create_repo(repo_id=f"{HF_ORG}/{DATASET_NAME}", repo_type="dataset", exist_ok=True).
  • upload_large_folder(folder_path="release/", repo_id=..., repo_type="dataset", ignore_patterns=["scripts/__pycache__/*"]).
  • Prints the resulting URL.

Step-by-step the user runs

  1. Create venv and install deps
    • python -m venv .venv && source .venv/bin/activate
    • pip install -r release/scripts/requirements.txt
  2. Generate the release artifacts
    • python release/scripts/prepare_release.py
    • Verify release/data/... shape and row counts in stdout.
  3. Edit release/README.md: fill in <HF_ORG>/<DATASET_NAME>, citation, license confirmation, and any per-config notes.
  4. Authenticate with HF: huggingface-cli login (paste a write-scoped token).
  5. Push:
    • python release/scripts/upload_to_hf.py --org <HF_ORG> --name <DATASET_NAME>
    • Or equivalently: huggingface-cli upload <HF_ORG>/<DATASET_NAME> release/ . --repo-type=dataset.
  6. On the HF dataset page, confirm the config/split selector shows all 7 splits and that the data viewer renders rows.

Things explicitly NOT done

  • No edits to any source JSON in data/data/ or data/evaluation/data_input_dir/.
  • No upload of audio files; users are pointed at upstream datasets for the audio.
  • Repo name kept as a placeholder until the user provides it.