embed listening samples: teacher vs both LR arms
Browse files- .gitattributes +3 -0
- README.md +16 -0
- samples/student_5.7B_lr-3e-5.wav +3 -0
- samples/student_5.7B_lr-6e-5.wav +3 -0
- samples/teacher_8B_reference.wav +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
samples/student_5.7B_lr-3e-5.wav filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
samples/student_5.7B_lr-6e-5.wav filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
samples/teacher_8B_reference.wav filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -18,6 +18,22 @@ tests (recognizable genre, structure, and vocals on held-out prompts) while comp
|
|
| 18 |
than the teacher** (single-row / CFG-free: the distillation targets are the CFG-guided
|
| 19 |
distributions, so guidance is baked in and no unconditional pass is needed).
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
| subfolder | LR | notes |
|
| 22 |
|---|---|---|
|
| 23 |
| `lr-3e-5/` | 3e-5 | conservative anneal arm |
|
|
|
|
| 18 |
than the teacher** (single-row / CFG-free: the distillation targets are the CFG-guided
|
| 19 |
distributions, so guidance is baked in and no unconditional pass is needed).
|
| 20 |
|
| 21 |
+
## Listen — same prompt, same seed, same renderer (mainstream arena rock, 40 s)
|
| 22 |
+
|
| 23 |
+
Only the composer differs. Teacher = stock 8B dual-row CFG path; students compose single-row, ~2x faster.
|
| 24 |
+
|
| 25 |
+
**Teacher (8B, reference):**
|
| 26 |
+
|
| 27 |
+
<audio controls src="https://huggingface.co/Mothersuperior/minimax-music3-composer-5.7b-distilled/resolve/main/samples/teacher_8B_reference.wav"></audio>
|
| 28 |
+
|
| 29 |
+
**Student 5.7B — lr 3e-5 arm:**
|
| 30 |
+
|
| 31 |
+
<audio controls src="https://huggingface.co/Mothersuperior/minimax-music3-composer-5.7b-distilled/resolve/main/samples/student_5.7B_lr-3e-5.wav"></audio>
|
| 32 |
+
|
| 33 |
+
**Student 5.7B — lr 6e-5 arm:**
|
| 34 |
+
|
| 35 |
+
<audio controls src="https://huggingface.co/Mothersuperior/minimax-music3-composer-5.7b-distilled/resolve/main/samples/student_5.7B_lr-6e-5.wav"></audio>
|
| 36 |
+
|
| 37 |
| subfolder | LR | notes |
|
| 38 |
|---|---|---|
|
| 39 |
| `lr-3e-5/` | 3e-5 | conservative anneal arm |
|
samples/student_5.7B_lr-3e-5.wav
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:97a358659b4453debd2b3df6ed6655400839cd3b4b069a1e13152b403df0b32f
|
| 3 |
+
size 7063596
|
samples/student_5.7B_lr-6e-5.wav
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d4b769ed9aa0a96f6695f6c471b6fcf795ec03b0c5049c0bebf4bffbeee2d037
|
| 3 |
+
size 7063596
|
samples/teacher_8B_reference.wav
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7095f7a3b40341d9a45c76c2453fab1492093f9456fa203019a3219b589875ba
|
| 3 |
+
size 7063596
|