train

name	train
description	Train a FineType CharCNN model with hardware auto-detection (Metal/CUDA/CPU)
user-invocable	true

Train a FineType Model

Run from the finetype repo root (~/github/noon-org/finetype/).

Quick Start

# Quick test (CPU, ~5 min)
./scripts/train.sh --samples 100 --size small --epochs 2

# Standard training
./scripts/train.sh --samples 1000 --size small --epochs 10

# Large model on M1 Mac (Metal auto-detected)
./scripts/train.sh --samples 5000 --size large --epochs 15 --seed 42

Architecture Presets

Preset	embed_dim	num_filters	hidden_dim
small	32	64	128
medium	64	128	256
large	128	256	512

Override individual params: --embed-dim 64 --num-filters 128 --hidden-dim 256

Hardware Detection

• macOS -> Metal GPU (Apple Silicon)
• Linux + NVIDIA -> CUDA GPU
• Otherwise -> CPU fallback

The script passes --features metal or --features cuda to Cargo automatically.

What It Does

1. Generate training data (finetype generate --samples N)
2. Build CLI with correct hardware features
3. Train CharCNN with progress display (epoch/loss/accuracy/ETA)

Output goes to models/char-cnn-vN/ (auto-incremented). Training log saved alongside.

All Flags

--samples N         Samples per type (default: 1000)
--size PRESET       small|medium|large (default: small)
--epochs N          Training epochs (default: 10)
--seed N            Random seed (default: 42)
--embed-dim N       Override embedding dimension
--num-filters N     Override CNN filters
--hidden-dim N      Override hidden layer dimension
--model-name NAME   Output dir name (default: auto char-cnn-vN)
--data FILE         Use existing NDJSON (skip generation)

After Training

./scripts/eval.sh --model models/char-cnn-vN    # Evaluate
./scripts/package.sh models/char-cnn-vN          # Package for distribution