| name | eval-lyrics |
| description | Audit and evaluate nano's lyric data and lyric conditioning. Use this skill when the user wants to analyze the lyric dataset, check transcribe/phonemize coverage or backfill progress, measure the Whisper-hallucination rate, see the vocal-gender or word-count distribution, decide whether the lyric filter should run, or evaluate whether a checkpoint actually sings the words. |
| allowed-tools | Read, Bash |
Evaluate lyrics โ data health and model intelligibility
Want the whole-corpus readiness view (all streams + scale + melody +
duration), not just lyrics? Use eval-training-data โ it wraps the same
audit and routes back here for the lyric detail below.
Two distinct questions live here. Route first:
| Question | Command |
|---|
| Data: how healthy is the lyric corpus right now? (coverage, hallucination rate, gender, word counts) | modal run scripts/lyrics_audit.py |
| Data: exactly what would the hallucination filter null? | modal run --detach diskrot/modal_filter_lyrics.py (dry-run report, prints to remote logs) |
| Model: does a checkpoint sing intelligible words? (WER) | modal run scripts/eval_lyric_wer.py โ details in the eval-checkpoint skill |
| Model: is lyric conditioning earning its compute? (val-loss ablation) | modal run scripts/eval_lyrics_ablation.py โ details in the eval-checkpoint skill |
The model-side rows are owned by eval-checkpoint; this skill owns the
data side.
The data audit โ scripts/lyrics_audit.py
One CPU container, read-only sweep of nano-tokens + the corpus listing.
Safe to run while transcribe/filter is in flight โ it's a snapshot of disk.
Reports per-stream coverage of the packed corpus (tags / structure / keys /
lyrics / phonemes), the lyric breakdown, gender + word-count distributions
among vocal-ready songs, and the top detected keys.
Reading the lyric breakdown
Entry semantics in lyrics/ (the audit reports all four buckets):
- null = transcribed, no vocals found โ trains as
<instrumental>
- has words, flagged = Whisper hallucination (a caption like "Thank you."
invented over instrumental audio) โ the filter will null it
- has words, clean = vocal-ready, the songs that teach singing
- ABSENT = never transcribed โ trains as
<unknown_vocals>
Expected rates from the 2026-06 full-corpus sweeps: ~48% instrumental(null),
~25% of with-words entries hallucinated, ~38% of transcribed songs
vocal-ready. A hallucination rate far above ~25% or a vocal-ready share far
below ~โ
deserves investigation before training.
Vocal gender now comes from the auto_tag captioner, not transcribe: the
audio-LLM writes a GENDER: tag into tags.json and the dataset reads that
first, falling back to the legacy F0 gender field in lyrics/ (only written
when transcribe runs with estimate_gender=True โ the Modal stage doesn't).
So a missing gender is backfilled by re-captioning (modal_auto_tag --redo),
NOT by re-running transcribe.
Acting on the audit
- Lyrics/phonemes coverage low โ transcribe backfill incomplete (check
modal app list for a running nano-transcribe app) or phonemize hasn't
run since the last transcribe.
- Hallucinated count > 0 โ run the filter, but only after the transcribe
fleet has fully finished โ its orchestrator holds shard contents in
memory and the next flush would clobber concurrent edits. Dry-run is
read-only and always safe;
--apply is the dangerous one.
- Required order: transcribe โ
modal run --detach diskrot/modal_filter_lyrics.py --apply โ optional forced alignment (modal run --detach diskrot/modal_align_lyrics.py --wave-id N --apply, sharpens word onsets
in place) โ modal run --detach diskrot/modal_phonemize.py โ re-audit โ
train. Phonemize after the filter, so junk never enters the phoneme store.
Prereqs
- The packed corpus must exist on
nano-tokens (the audit measures coverage
of the packed set; see the add-songs skill for the pipeline).
- Model-side evals need a checkpoint โ see eval-checkpoint for flags,
baselines, and interpretation.
Next step
Data healthy and fully phonemized โ train-model. Checkpoint in hand โ
eval-checkpoint (WER row) to track intelligibility across steps.
