lora-train

What problem does it solve?

It solves the problem of turning a prepared and captioned dataset into a trainable LoRA model without manually wiring backend-specific training steps.

Core Features & Use Cases

• Backend-agnostic LoRA training orchestration: Runs training via kohya_ss, sd-scripts, or OneTrainer while keeping the calling interface consistent.
• Dataset-to-model production: Consumes an absolute dataset path and writes the resulting .safetensors plus run metadata into an explicit output folder.
• Configurable budgets and overrides: Uses budget_tier presets (low/standard/high) with arbitrary config overrides for rank, alpha, learning rate, steps, and resolution.
• Deterministic job lifecycle: Builds backend config, submits the job, polls progress, downloads/moves outputs, and returns a candidate status without promotion.

Quick Start

Use lora-train to train a LoRA by providing the dataset_path, output_path, model_name, trigger_token, lora_class, base_model, and backend, then verify the returned candidate model path and run metadata.