---

name: evaluation description: "Compute and interpret SketchCode BLEU scores for single or batch GUI DSL files." disable-model-invocation: true metadata: disco-role: operating license: MIT

SketchCode evaluation

Use this sub-skill when the task is to evaluate SketchCode-style .gui output, compute BLEU for one generated GUI, compare original and predicted GUI folders, explain button-color normalization, or debug NLTK BLEU warnings.

Start here

1. For commands and workflow choices, read references/evaluation-workflow.md.
2. For the distilled Evaluator behavior, normalization rules, prediction trimming, and batch pairing contract, read references/api-reference.md.
3. For skipped files, low BLEU, NLTK warnings, short-sequence behavior, or surprising button-color results, read references/troubleshooting.md.
4. For a self-contained smoke check or file-based local score without depending on original source scripts, run scripts/evaluate_tiny_gui_bleu.py.

Route elsewhere

• If the user still needs to generate predicted .gui files from sketches or HTML conversion outputs, route to ../conversion-inference/SKILL.md.
• If the user is preparing data, training, fine-tuning, or validating paired training examples, route to ../training-data/SKILL.md.

This sub-skill distills the SketchCode evaluation behavior into bundled references and a safe helper. Do not require runtime access to the original repository scripts just to understand normalization, pairing, or BLEU interpretation.