lora-evaluate

What problem does it solve?

This Skill helps you catch LoRA quality issues early by generating a controlled test grid and scoring consistency, fidelity, overfitting, and visual artefacts before promotion.

Core Features & Use Cases

• LoRA test-grid generation: Produces a structured set of evaluation images from a freshly trained LoRA using configurable prompts and weight sweeps.
• Multi-axis automated scoring: Evaluates subject consistency, prompt fidelity, overfitting risk, and artefacts, then outputs a pass/review/fail verdict with evidence.
• Negative-control leakage detection: Includes a prompt that should not mention the subject to detect concept leakage.
• Pipeline-safe, read-only evaluation: Does not train, promote, delete, or publish images—returns an advisory result for human decision-making.

Quick Start

Ask lora-evaluate to evaluate the newly trained LoRA by providing the model_path, run_metadata_path, and output_path so it can write the grid and evaluation report.