mm3-captioning

Write a Structured Caption with three plain-text sections in order: Global Metadata, Vocal Details, and Arrangement. Keep it under the 5,000-token budget shared with lyrics, and never quote or paraphrase lyric text inside the caption.

Short one-line captions produce high take-variance and genre drift in MM3. Detailed three-section Structured Captions are the adherence lever, and the prose vocabulary must match the target genre family's dialect in the upstream templates, not just name the genre.

Use plain-text section labels like Global Metadata, Vocal Details, and Arrangement with no markdown syntax. The reference templates and the tokenizer's caption slot treat the caption as an opaque string, and the trained format uses plain labels.

Mechanical restructuring of ACE captions consistently underperforms hand-written Structured Captions in ear-judged tests, even when the content is rich. The reliable path is captioning from the audio directly in MM3 format, for example with ace-caption --mode mm3.

Captioning models are unreliable on tempo and key, so rebuild the Basic Attributes line deterministically from measured values such as Essentia bpm, keyscale, and time signature. Preserve the model-written genre clause, since genre judgment is where the model performs better.

Spectral proxies like flatness, centroid, and intro ratio repeatedly mislead when evaluating caption changes, because genre arrival itself shifts those numbers. Caption adherence for MM3 must be judged by ear across multiple seeds.