sbob-datasets / README.md
weystrom's picture
Upload README.md with huggingface_hub
d499a98 verified
|
Raw History Blame Contribute Delete
1.83 kB
# sbob-datasets — SpongeBob LoRA + TTS pipeline data
Everything needed to rebuild any stage of the sbob project without the original
episode files. **Private, personal research only** — derived from © Nickelodeon material.
| dir | contents |
|---|---|
| `corpus/` | `lines.jsonl` — 129k speaker-attributed transcript lines from 1043 episodes (the scraped source corpus); `episodes.jsonl` index |
| `text_datasets/` | per-character LoRA train/val jsonls (10 chars) + `_ensemble/` (103 speakers, 77k records) — completion format, episode-level split; **the ensemble adapter trained on this beat every per-char adapter** (see `corpus_eval/REPORT.md`) |
| `tts_ft_v3/` | trio voice-FT jsonls, round-3 recipe: fixed slots S1/S2/S3, singles-only, slow-Patrick filter |
| `tts_clips/` | the exact 24 kHz mono training clips (cut from demucs vocals at attribution timings) |
| `media_vocals_flac/` | demucs (htdemucs) vocal stems of all 41 S01 episodes, lossless FLAC, 48k mono — re-cut clips/refs from these |
| `refs_v2/` | frozen cloning refs (vocals-domain) + alternates for the trio — **use these with the v3 checkpoints** |
| `refs_v1/` | original ref pools + manifests (raw-audio domain, ref-grade clips, all attributed chars) |
| `attribution/` | SRT × transcript alignments per episode (speaker, start/end, text, score) — the timing ground truth |
| `eval_probes/` | the frozen probe/teaser/scene jsonls used for every voice A/B |
| `outlines/` | episode "director" scripts: The Laugh Tax, quantum podcast (+ trio cut) |
| `corpus_eval/` | text-tune comparison: 12-config × 11-val-set matrix, gold-referenced samples, REPORT.md |
Checkpoints trained on this data: `weystrom/sbob-tts-ckpts`. Tooling: the `sbob` repo
(`scrape.py`, `build_dataset.py`, `tts/attribute_voices.py`, `tts/build_tts_ft_dataset.py`).