mirror of
https://github.com/Nighthawk42/soprano-factory.git
synced 2026-08-30 04:30:21 +00:00
- Move hyperparameters from hardcoded script values to `config.yaml` - Replace pip requirements with `pyproject.toml` and `uv` support (CUDA 12.6) - Refactor all scripts to use `pathlib` for robust path handling - Optimize `generate_dataset.py` with GPU acceleration - Register quantizer constants as buffers for proper device mapping - Update README with new installation and usage instructions
20 lines
412 B
YAML
20 lines
412 B
YAML
# Hardware & Reproducibility
|
|
device: "cuda:0"
|
|
seed: 1337
|
|
|
|
# Learning Rate Schedule
|
|
max_lr: 5.0e-4 # Use 5.0e-4 (with dot) to ensure it loads as a float
|
|
warmup_ratio: 0.1
|
|
cooldown_ratio: 0.1
|
|
|
|
# Training Dynamics
|
|
batch_size: 4
|
|
grad_accum_steps: 1
|
|
seq_len: 1024
|
|
max_steps: 10000
|
|
val_freq: 250
|
|
|
|
# Optimizer & Model Config
|
|
betas: [0.9, 0.95]
|
|
weight_decay: 0.1
|
|
text_factor: 0.0 # Increase to train on text inputs |