choihy/LMD_tokens_and_pr

LMD Simple-Pack Shards: Download and Unpack

This document explains how to download the simple-packed LMD training shards from Hugging Face and restore them back into the raw grouped-sample directory tree expected by the local LMD loader.

The unpack step uses lmd_training_shards_unpack.py. It can be run either:

• from this repository: pretrain/data/lmd_training_shards_unpack.py
• from the downloaded Hugging Face dataset: scripts/lmd_training_shards_unpack.py

What Is In The Hugging Face Dataset

The uploaded dataset is expected to follow this layout:

<download_root>/
  train/
    piano/piano_train_shard_00000.tar
    bass/bass_train_shard_00000.tar
    ...
  val/
    ...
  eval/
    ...
  manifests/
    train/piano.jsonl
    train/bass.jsonl
    val/...
    eval/...
  scripts/
    lmd_training_shards_unpack.py
  repack_summary.json
  SHA256SUMS                # optional

In simple tar mode, each tar shard stores the original event_segment/*.pt and pianoroll_segment/*.pt.gz files by their source-root-relative paths.

The matching manifests/<split>/<instrument>.jsonl files are required for unpacking. Do not download only the tar files without the manifests/ directory.

1. Download From Hugging Face

Replace YOUR_HF_DATASET_REPO with your dataset repo id on Hugging Face.

Option A: Python (snapshot_download)

pip install huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YOUR_HF_DATASET_REPO",
    repo_type="dataset",
    local_dir="/path/to/lmd_simple_pack",
)

Option B: CLI (hf download)

pip install "huggingface_hub[cli]"
hf download YOUR_HF_DATASET_REPO \
  --repo-type dataset \
  --local-dir /path/to/lmd_simple_pack

Download Only A Subset

If you only need specific splits or instruments, make sure to download both:

• the matching tar shards under train/, val/, or eval/
• the matching manifest JSONL files under manifests/

Example with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YOUR_HF_DATASET_REPO",
    repo_type="dataset",
    local_dir="/path/to/lmd_simple_pack",
    allow_patterns=[
        "train/piano/*.tar",
        "manifests/train/piano.jsonl",
        "scripts/lmd_training_shards_unpack.py",
        "repack_summary.json",
        "SHA256SUMS",
    ],
)

If the dataset is private or gated, log in first:

hf auth login

2. Optional: Verify Checksums

If the repo includes SHA256SUMS, you can verify the downloaded shards:

cd /path/to/lmd_simple_pack
sha256sum -c SHA256SUMS

3. Unpack The Shards

You can use either the local repository script or the copy downloaded from Hugging Face.

Option A: Use The Repository Script

cd /path/to/think-before-you-play
python pretrain/data/lmd_training_shards_unpack.py \
  --source-root /path/to/lmd_simple_pack \
  --output-root /path/to/lmd_unpacked

Option B: Use The Downloaded Hugging Face Script

cd /path/to/lmd_simple_pack
python scripts/lmd_training_shards_unpack.py \
  --source-root /path/to/lmd_simple_pack \
  --output-root /path/to/lmd_unpacked

This script:

• finds all *_shard_00000.tar-style files under source-root
• extracts the original event and pianoroll tensors
• rebuilds _sample_index/<split>/<instrument>.jsonl
• writes unpack_summary.json

The output directory must be empty. If you want to replace an existing output directory, add:

python pretrain/data/lmd_training_shards_unpack.py \
  --source-root /path/to/lmd_simple_pack \
  --output-root /path/to/lmd_unpacked \
  --overwrite-output

4. Unpack Only Selected Splits Or Instruments

You can filter during unpack:

python pretrain/data/lmd_training_shards_unpack.py \
  --source-root /path/to/lmd_simple_pack \
  --output-root /path/to/lmd_unpacked \
  --splits train val \
  --target-instruments piano bass

5. Output Layout

After unpacking, the directory will look like:

<output_root>/
  train/<bucket>/<md5>/event_segment/*.pt
  train/<bucket>/<md5>/pianoroll_segment/*.pt.gz
  val/<bucket>/<md5>/...
  eval/<bucket>/<md5>/...
  _sample_index/
    train/piano.jsonl
    train/bass.jsonl
    val/...
    eval/...
  unpack_summary.json

This unpacked tree matches the grouped LMD layout used by dataset.loaders.lmd_token_pr_load, so the raw loader can read it immediately without rebuilding the manifest again.

Notes

• Run the unpack script in an environment that has torch installed.
• You can run either python pretrain/data/lmd_training_shards_unpack.py ... from this repo or python scripts/lmd_training_shards_unpack.py ... from the downloaded Hugging Face dataset.
• If you downloaded only part of the dataset, the unpack filters should match the shards and manifest files you kept locally.
• For simple tar mode, missing manifest JSONL files will cause unpacking to fail.

References

• Hugging Face download guide: https://huggingface.co/docs/huggingface_hub/guides/download
• Hugging Face CLI reference for hf download: https://huggingface.co/docs/huggingface_hub/package_reference/cli