Update: best-rank release (lora_r=32) from rank sweep

README.md

@@ -18,7 +18,7 @@ pipeline_tag: text-generation
 
 # TheArtist Music Transformer — LoRA Adapter (Blues)
 
-**LoRA adapter that conditions the F1 base (PearlLeeStudio/TheArtist-MusicTransformer-ft-pop80) toward blues chord progressions.** One of eleven per-genre adapters released alongside the paper *Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation* (Lee, 2026).
+**LoRA adapter that conditions the F1 base (PearlLeeStudio/TheArtist-MusicTransformer-ft-pop80) toward blues chord progressions.** One of eleven per-genre adapters released alongside the paper *Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation* (Lee, 2026). This release is the **best-rank** snapshot from a 5-point rank sweep (`r ∈ {4, 8, 16, 32, 64}`); see §Rank sweep below for the full table and selection criterion.
 
 ## Adapter summary
 
@@ -26,35 +26,63 @@ pipeline_tag: text-generation
 |---|---|
 | Base model | `PearlLeeStudio/TheArtist-MusicTransformer-ft-pop80` (F1, 25.6M params) |
 | Adapter type | LoRA (Q/K/V projections) |
-| LoRA rank | 8 |
-| LoRA alpha | 16 |
+| LoRA rank | **32** |
+| LoRA alpha | 64 |
 | LoRA dropout | 0.05 |
 | Target modules | `w_q`, `w_k`, `w_v` |
-| Trainable parameters | ~200K (~0.78% of base) |
-| Adapter file size | ~800 KB |
+| Trainable parameters | ~794,624 (~3.06% of base) |
+| Adapter file size | ~3.0 MB |
 | Base vocabulary | 351 tokens (jazz/pop) |
-| Vocabulary extension | +13 genre tokens (`embedding_extension.pt`) |
+| Vocabulary extension | +8 genre tokens (`embedding_extension.pt`) |
 | Training epochs | 8 |
 
 ## Training data
 
 9,944 chord-progression sequences in the **blues** subset of the [Chordonomicon dataset](https://huggingface.co/datasets/ailsntua/Chordonomicon). Chordonomicon is licensed CC BY-NC 4.0; see the dataset card for full terms.
 
+Training data preparation: per-genre extraction by `ai/training/extract_genre_subsets.py` (`main_genre` match OR `genres` column substring match; first-match-only to avoid double-counting). 12-key augmentation × 80/10/10 song-level train/val/test split. Same pipeline + same val split as the F1 baseline so the LoRA-vs-base comparison is apples-to-apples.
+
 ## Genre character
 
 12-bar and modified blues forms
 
+## Rank sweep
+
+The released adapter is the **best-rank** snapshot from training the same LoRA recipe at five different ranks. Every cell uses the same F1 base, same val split, same `evaluate()` call, and the same `[GENRE:none]`-initialized embedding extension — only `lora_r` (and `lora_alpha = 2 × lora_r`) changes. Numbers are validation-set token-level metrics (no key augmentation).
+
+| Rank | val_loss | val_top1 (%) | val_top5 (%) | Δtop1 vs F1 |
+|---|---:|---:|---:|---:|
+| r=4 | 0.5647 | 84.16 | 96.78 | +2.46 |
+| r=8 | 0.5655 | 84.13 | 96.77 | +2.43 |
+| r=16 | 0.5640 | 84.16 | 96.76 | +2.46 |
+| **r=32** | 0.5632 | 85.05 | 96.78 | +3.35 ← selected |
+| r=64 | 0.5652 | 84.14 | 95.86 | +2.44 |
+
+**Selection criterion**: minimum validation cross-entropy loss; `val_top1` as tiebreaker. `val_loss` is what the training loop optimizes and what selects each rank's `best.pt` epoch, so using it for cross-rank selection keeps consistency with how each individual checkpoint was chosen.
+
+Full 11-genre × 5-rank sweep + full-FT anchor table: `ai/results/lora_rank_sweep.md` in the repo.
+
 ## Evaluation
 
-Validation token-level metrics on the genre-specific val split (994 sequences, no key augmentation). The F1 base column uses the same val split, same dataloader, and the same `[GENRE:none]`-initialized embedding-extension setup as the LoRA run — only the LoRA parameters and the trained embedding rows differ. The LoRA snapshot is the best-validation epoch (3/8).
+Validation token-level metrics on the genre-specific val split (994 sequences, no key augmentation). The F1 base column uses the same val split, same dataloader, and the same `[GENRE:none]`-initialized embedding-extension setup as the LoRA run — only the LoRA parameters and the trained embedding rows differ.
 
 | Metric | F1 base alone | F1 + this LoRA | Δ |
 |---|---:|---:|---:|
-| Top-1 accuracy (%) | 81.7 | **84.1** | **+2.40** |
-| Top-5 accuracy (%) | 94.8 | 96.8 | +2.00 |
-| Cross-entropy loss | 0.8137 | 0.5655 | -0.2482 |
+| Top-1 accuracy (%) | 81.70 | **85.05** | **+3.35** |
+| Top-5 accuracy (%) | 94.80 | 96.78 | +1.98 |
+| Cross-entropy loss | 0.8137 | 0.5632 | -0.2505 |
+
+Source: `ai/results/f1_per_genre_baseline.csv` + `ai/results/lora_rank_sweep.csv`. Higher top-1/top-5 and lower loss are better.
+
+## Real-song eval
 
-Source: `ai/results/f1_per_genre_baseline.csv` + `ai/logs/ft_f1_lora_blues_*.log`. Higher top-1/top-5 and lower loss are better. The 11-adapter comparison and the genre-distance-vs-gain pattern are reported in the 2026 workshop paper.
+Mean validation top-1/top-5/cross-entropy on **10 held-out real blues songs** from `ai/data/eval_real_songs.jsonl` (held-out from `ai/data/splits/{val,test}.jsonl`, see [docs/EVAL.md](https://github.com/JinJuLee/PearlLeeStudio_TheArtist/blob/main/docs/EVAL.md) for dataset composition + methodology). Teacher-forced eval — same `evaluate()` call as the full-val rank-sweep eval above, just narrowed to a curated 10-song subset.
+
+| Model | Top-1 (%) | Top-5 (%) | val_loss |
+|---|---:|---:|---:|
+| F1 base alone | 25.00 | 50.00 | 6.8125 |
+| **F1 + this LoRA** | **75.00** | 100.00 | 3.0176 |
+| Δ | +50.00 | +50.00 | -3.7949 |
 
 ## License and use
 

adapter_config.json

@@ -19,7 +19,7 @@
   "layers_pattern": null,
   "layers_to_transform": null,
   "loftq_config": {},
-  "lora_alpha": 16,
+  "lora_alpha": 64,
   "lora_bias": false,
   "lora_dropout": 0.05,
   "lora_ga_config": null,
@@ -29,13 +29,13 @@
   "peft_type": "LORA",
   "peft_version": "0.19.1",
   "qalora_group_size": 16,
-  "r": 8,
+  "r": 32,
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
+    "w_v",
     "w_q",
-    "w_k",
-    "w_v"
+    "w_k"
   ],
   "target_parameters": null,
   "task_type": null,

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:878823df46e438b8d9373e1d41883f18dc7c95f23b8d1216d658abe7ea43cfeb
-size 791928
+oid sha256:638dc66cc20306d7e3800b35920da46843c93db778d5fc671baa77f202559a86
+size 3151352

config.json

@@ -1,5 +1,5 @@
 {
-  "run_name": "ft_f1_lora_blues",
+  "run_name": "ft_f1_lora_blues_r32",
   "base_checkpoint": "checkpoints/ft_jazz_pop80/best.pt",
   "target_genre": "blues",
   "sources": [
@@ -11,8 +11,8 @@
   "n_layers": 8,
   "max_seq_len": 256,
   "dropout": 0.1,
-  "lora_r": 8,
-  "lora_alpha": 16,
+  "lora_r": 32,
+  "lora_alpha": 64,
   "lora_dropout": 0.05,
   "lora_target_modules": [
     "w_q",

embedding_extension.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75733c6ff47dd0e9f35d7196e1636281e2fc6a6c9e7af363664334c19dc7bd1f
+oid sha256:1f7b5f7fab33a785d45965a74e067a956749bc6be6ee554ef13f9a566f943a28
 size 1472800