Add Musicroll-50M model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: mlx
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
tags:
|
| 5 |
+
- music
|
| 6 |
+
- piano-roll
|
| 7 |
+
- causal-lm
|
| 8 |
+
- qwen3
|
| 9 |
+
- safetensors
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Musicroll-50M
|
| 13 |
+
|
| 14 |
+
Musicroll-50M is a small Qwen3-style causal language model over a 259-token
|
| 15 |
+
byte-level music vocabulary. It was used as the MusicRoll negative-control model
|
| 16 |
+
in the replication artifact for "Many Next-Token Predictors are In-Context
|
| 17 |
+
Learners".
|
| 18 |
+
|
| 19 |
+
The checkpoint is provided as `model.safetensors` with a compact tokenizer:
|
| 20 |
+
|
| 21 |
+
- vocabulary size: 259
|
| 22 |
+
- hidden size: 512
|
| 23 |
+
- layers: 16
|
| 24 |
+
- attention heads: 4
|
| 25 |
+
- key/value heads: 2
|
| 26 |
+
- context length: 4096
|
| 27 |
+
- BOS/EOS/PAD token ids: 256/257/258
|
| 28 |
+
|
| 29 |
+
The bitstring replication harness encodes tasks as piano-roll-like text and
|
| 30 |
+
loads this checkpoint through the local MLX model loader.
|
| 31 |
+
|
| 32 |
+
## Intended Use
|
| 33 |
+
|
| 34 |
+
This upload is primarily for reproducing the MusicRoll negative-control runs in
|
| 35 |
+
the ICLManyReplication artifact. General music-generation behavior has not been
|
| 36 |
+
characterized here.
|