Files
soprano-factory/config.yaml
T
Nighthawk 9b7dd275c7 Refactored training pipeline to use YAML config and uv
- 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
2026-01-18 01:13:38 -05:00

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