TheMindExpansionNetwork/nps-liminal-soundscapes-v0-2

nps-liminal-soundscapes-v0-2

Fork + repair of `Sonic-Forage/nps-liminal-soundscapes-v0-1` (org-create permissions on our token are restricted, so the improved version lives under TheMindExpansionNetwork; upstream keeps the v0-1 original untouched).

Sonic-Forage Stable Audio soundscape training dataset built from verified National Park Service public-domain sound recordings (Rocky Mountain National Park).

v0-2 is a re-release of v0-1 with layout + caption fixes (see Changelog below).

Contents

• —118 train / 10 validation clips, 48 kHz stereo WAV, loudness-normalized (ffmpeg loudnorm), ~8s each
• —Layout: data/train/*.wav + data/validation/*.wav, each with its own metadata.csv (Hugging Face audiofolder-compatible; verified with datasets 5.x)
• —Captions: public-domain natural soundscape identity, explicitly no music / no drums / no human voice

Caption design (improved in v0-2)

v0-1 had only 12 caption strings shared across 128 clips. v0-2 enriches every clip with:

• —Species identity for all 89 bird clips (e.g. "lincoln's sparrow song, a long delicate buzzy warble")
• —Habitat context parsed from source titles/URLs (dawn, open meadow, forest, near moving water…)
• —Measured audio characteristics (RMS loudness bucket + dynamics variance bucket from per-clip analysis)
• —Spectral texture class for ambient clips (low/mid/bright)

→ 48 distinct caption templates across 118 train clips.

Columns (metadata.csv)

file_name, text, duration, id, source_page, source_clip_url, source_title, source_collection, license, credit, tags, sample_rate, channels, split, sha256, rms_db

Every row carries the NPS source page URL, the exact source audio URL, and a SHA-256 of the clip file (all 128 verified on rebuild).

Target model

Stable Audio / Stable Audio 3 LoRA first pass.

Load

from datasets import load_dataset
ds = load_dataset("Sonic-Forage/nps-liminal-soundscapes-v0-2")
# ds["train"] -> 118, ds["validation"] -> 10, columns incl. "audio" + "text"

Credit & license

All recordings are public domain per the NPS source pages. Please credit: National Park Service. Full per-clip source URLs + SHA-256 hashes in sources/metadata.full.jsonl; the v0-1 original metadata is preserved at sources/metadata.v0-1.original.jsonl; license receipts in LICENSE_RECEIPT.md and sources/.

Changelog

• —v0-2 (2026-08-14): HF data/{train,validation}/ audiofolder layout (v0-1's dual metadata.csv+metadata.jsonl at root + deprecated README config no longer loads in datasets ≥4.x); caption enrichment 12→48 unique with species/habitat/acoustics; splits now real (v0-1 loaded everything as train despite a split column); README accuracy (v0-1 referenced Yellowstone license pages — all clips are ROMO); rms_db column added; provenance archived.
• —v0-1 (2026-05-26): initial 128-clip build.