PearlLeeStudio/TheArtist-MusicTransformer-lora-bossa

Exact trainable totals on the data line (760,832-style counts replace 761K-style rounding)

Data-line precision: training counts are windowed sequences; unique song counts added (e.g. rock 48,910 sequences from 42,781 songs)

Card corrections: adapter parameter accounting (24,576*r file + 367,616 embedding/output = trainable set, % of the 25,665,536 full-FT set); classical split provenance; released-checkpoint epoch note; fix a section reference

Rewrite model card: clean standard form (usage / eval / training data / license)

docs: standardize link-block (GitHub/@StudioPearlLee), Phase-0, real-song-eval format

docs: R9 cosmetic polish (formatting/wording/citation; no number changes)

docs: real_song_eval.csv 18-model count + frontmatter normalize

docs: real-song-eval Δ cells made internally consistent; F5 wording; v2 18-model count; classical/Bach

Card erratum: released F1 weights coincide with Phase-0 baseline (selection artifact)

Update model card: per-genre LoRA eval table

Add pipeline tag and links to paper/code/demo (#1)

docs: fill arXiv2 ID (2606.07334) in citation (companion LoRA paper now on arXiv)

Add generated audio samples (chord-only + adapter/base output)

Add demo video link (youtu.be/iSDj4VK9nKc)

Update model card: per-genre LoRA eval table

Update: best-rank release (lora_r=16) from rank sweep

Update: best-rank release (lora_r=16) from rank sweep

Update model card: per-genre LoRA eval table

Fix adapter_config.json: set base_model_name_or_path + task_type (HF warning)

Update: best-rank release (lora_r=16) from rank sweep

Update: best-rank release (lora_r=16) from rank sweep

Update model card: add per-genre LoRA evaluation table (workshop paper)

Initial release: TheArtist genre LoRA adapter (F1 base)

initial commit