LDM with subdivision, M2 at 3.6% training size

This is the M2 checkpoint selected by saved validation loss at the 3.6% training size in the LDM capacity experiment. It uses 4 layers, embedding width 192, 4 attention heads, and a 6,656-token context. Its normal training subset contains 9,830 cycles (16.15 hours). The exported checkpoint is from step 20,000.

Held-out test results

The saved normal test split contains 17 files, 34,184 cycles, and 56.10 hours. Evaluation used the archived LDM scorer on an NVIDIA A100 SXM 80 GB.

Metric Value
Exact HIT accuracy, including silence 71.471%
Overall token accuracy 81.003%
Skeleton HIT accuracy 94.133%
Test cross-entropy 0.395484
Saved checkpoint validation cross-entropy 0.395874

Usage

This snapshot uses atomic rhythm tokens and five-line .derbake files. It was built from a selected training checkpoint. trust_remote_code=True executes the model/processor code from the selected repository revision.

from transformers import AutoModel, AutoProcessor

repo_id = "Appodoom/ldm-with-subd-m2-3-6pct"
revision = "main"  # Replace with the verified release commit SHA for reproducibility.
processor = AutoProcessor.from_pretrained(repo_id, revision=revision, trust_remote_code=True)
model = AutoModel.from_pretrained(repo_id, revision=revision, trust_remote_code=True).eval()
inputs = processor(derbake_path="prompt.derbake", return_tensors="pt")
ids = model.generate(input_ids=inputs["input_ids"].to(next(model.parameters()).device),
                     num_cycles=10, max_new_tokens=4096)
rhythm = processor.decode_generation(ids[0], expected_num_cycles=10)
processor.save_derbake(rhythm, "generated.derbake")

The call returns only newly accepted cycles. Invalid sampled cycles are skipped with a Python warning; use return_generation_report=True for details. A token cap reached before ten valid cycles raises an error and produces no output file. The writer uses 120 BPM, amplitude 1.5, and a blank skeleton line. The model supports batch size one and at most 16 beats per cycle.

Downloads last month
-
Safetensors
Model size
1.2M params
Tensor type
I64
路
F32
路
BOOL
路
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support