lora-version

What problem does it solve?

It solves the problem of turning a finished LoRA (trained and evaluated) into a properly named, registry-backed, shareable, and auditable version without losing lineage to the dataset, config, and evaluation.

Core Features & Use Cases

• Resolves final LoRA identifiers using a chosen version strategy (semver, date, or increment) to avoid naming ambiguity.
• Creates an append-mostly registry entry as a portable denormalized record while preserving extra fields and preventing accidental overwrites.
• Consolidates and co-locates run artifacts by renaming/aligning .safetensors and sibling JSON files and generating a human-readable {final_id}.run.md report.
• Use case: after running lora-train and lora-evaluate for a candidate LoRA, use this skill to register the final version and produce a complete audit trail from training inputs to evaluation results.

Quick Start

Ask the system to version the evaluated LoRA at the absolute model_path using run_metadata_path, evaluation_path, and the provided registry_path, with an increment strategy and candidate status.