Mllm Eval

MLLM-Eval Skill

Purpose

This skill makes an LLM / MLLM clinical evaluation defensible: a real adjudicated reference standard, faithfulness measured not assumed, clinical-efficacy metrics beyond n-gram overlap, a pretraining- contamination check, prompt-sensitivity disclosed, and a reader study where text is generated. It is model-agnostic — every check applies to a closed API and to open weights — and read-only (an advisory design/audit skill): it audits the evaluation design and specifies and routes the clinical-efficacy metrics (RadGraph-F1 / CheXbert-F1 via their published extractors) rather than running the model or computing the metrics itself.

It is the LLM/MLLM evaluation-design counterpart in the lane — an auditor that hands the specified metrics to their extractors and /analyze-stats, parallel to how /model-assessment audits an imaging model's design and computes its metrics. The reviewer-side audit of a finished manuscript uses the mllm_evaluation.md (ME0–ME8) probe via /self-review and /peer-review; this skill is the author-side harness design. It routes the reader study to /design-ai-benchmarking, the sizing to /calc-sample-size, and TRIPOD-LLM / MI-CLEAR-LLM compliance to /check-reporting.

When to use

• You are designing or auditing an evaluation of an LLM/MLLM on a clinical task and want it to cover the axes a reviewer will check (reference standard, faithfulness, contamination, prompt sensitivity, reader study).

When NOT to use

• AI-vs-human-expert benchmark with a rated rubric → /design-ai-benchmarking.
• Imaging prediction/segmentation model → /model-assessment.
• Image-to-image generative model → the image_synthesis probe.
• Training / serving the LLM → out of scope.
• Item-level TRIPOD-LLM / MI-CLEAR-LLM audit of a finished manuscript → /check-reporting.

Workflow

Phase 1 — Pin the task, model, comparator, decoding (ME0)

State the task (report generation / VQA / extraction-classification), the exact model + version/date (closed API or open-weights id), the decoding settings (temperature, seed, max tokens), and what the outputs are scored against.

Phase 2 — Reference standard + metrics (ME1, ME2)

Require an adjudicated expert reference (not a single unverified report or a model-derived label). For report generation, report a clinical-efficacy metric — RadGraph-F1 (Jain et al., NeurIPS 2021) or CheXbert-F1 (Smit et al., 2020), or the composite RadCliQ (Yu et al., Patterns 2023) — alongside any BLEU/ROUGE, with CIs. For VQA/classification, state the answer-matching rule and report per-class sensitivity/specificity (or precision/recall/F1) and PPV at the real prevalence, with CIs; accuracy only alongside them — at 2% prevalence, answering "negative" every time scores 98% accuracy.

Phase 3 — Faithfulness + contamination (ME3, ME4)

Add an atomic-fact faithfulness measure + a false-premise / abstention probe (MedVH, Med-HALT) — report a hallucination rate, not just accuracy. For any public benchmark (VQA-RAD, SLAKE, MIMIC-CXR- derived, MedQA), add a contamination statement: training cutoff vs benchmark release, a held-out / post-cutoff set, or a contamination probe.

Phase 4 — Prompt sensitivity + reader study (ME5, ME7)

Disclose the exact prompt(s), temperature/seed, ≥ 3 runs with variance, and a prompt-robustness check. For a deployment/utility claim, design a blinded reader study with an error taxonomy (route the rubric/IRR to /design-ai-benchmarking, ICC/κ to /analyze-stats, sizing to /calc-sample-size).

Phase 5 — Gate the plan (deterministic)

Declare the axes in eval_manifest.json (copy ${CLAUDE_SKILL_DIR}/templates/eval_manifest.json; fields and allowed values in references/eval_manifest_schema.md), then:

python3 ${CLAUDE_SKILL_DIR}/scripts/check_mllm_eval_completeness.py \
  --manifest eval_manifest.json --strict

Use "none" for an axis not done and "other:<description>" for a method not listed; any other value exits 2. --plan plan.md --task report_generation|vqa|classification still runs the older keyword check on prose (see Known limits). NGRAM_ONLY / FAITHFULNESS_MISSING / REFERENCE_STANDARD_MISSING / CONTAMINATION_UNADDRESSED / READER_STUDY_MISSING must be resolved. A classification manifest always reports CLASSIFICATION_METRICS_NOT_ASSESSED (Minor): the manifest has no metric field for it, so check per-class sensitivity/specificity and PPV at the real prevalence by eye.