Began training on custom dataset for next note prediction. Using LightGBM as model. Trained on all instruments in order to teach the model music theory. Will finetune on guitar specifically after base training reaches acceptable level.
Did research and choose the options I will run tests on for different model options. Features are still up to testing and I will change those more than the model options
Small test result
Progress is on schedule. Plan to finish the larger training and analyze results. Will experiment with many different optimizations and features in order to increase accuracy. Should run many different ablations next week to test accuracy and inference latency.
Ran on the AMD GPUs in the 1305 clusters. Some stats
| GPU benchmark | matmul throughput | bf16 122 TFLOP/s, fp32 1.2 TFLOP/s (~100x gap) |
| Test suite | 16 unit tests | all pass on-box |
| Overfit test | 1,000 fixed windows, 3,000 steps | loss 18.1 → 0.012; guitar/bass/drum head accuracy 99.8%+ |
| Training slice | 18,284 songs, 6,000 steps, 197M tokens (~1 epoch of a 5% slice) | held-out loss 1.79, next-note top-1 67.5% (clean), 47.6% (±30ms jitter) |
Memory / throughput
Peak VRAM: 7.6 GB of 32 GB (24%)
Step time: 224 ms for batch 32 × 1024 tokens
Training throughput: ~146,000 tokens/s
Model size: 25.9M params (6-layer decoder, RoPE + SwiGLU)
bf16 check: every matmul in forward+backward confirmed bf16, zero fp32 fallback