testland/llm-eval-anti-patterns

Audits an existing LLM evaluation suite for eight methodology errors that make its numbers untrustworthy: too few cases per capability, single-provider lock-in, exact-match assertions on open-ended output, no semantic-similarity check on paraphrase-tolerant output, no baseline comparison in CI, no cost or latency ceiling, unpinned model identifiers, and no adversarial coverage. Supplies harness-neutral detection cues, the reason each error invalidates the result, a concrete fix, a Critical/Warning/Info severity scheme, and a findings-table output shape. Covers validating an LLM judge against human labels before its verdicts count as evidence. Use when an eval suite already exists and its pass rate is about to gate a release, a model swap, or a prompt change, and nobody has audited how the suite itself was built.

Quality

94%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Academic citations

Name: testland/llm-eval-anti-patterns
Rating: 75.2 (1 reviews)
Author: testland

Full references for the justifications in SKILL.md. Each parenthetical author-year token in the body resolves here.

(Miller 2024) Miller, Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations, arXiv:2411.00640, https://arxiv.org/abs/2411.00640. Evals are experiments; a suite's questions are a sample, so scores carry sampling error that must be quantified. A single run of a stochastic model is one draw. Running both variants on the same questions (paired structure) measures the difference far more precisely than comparing two independent scores.
(Magar and Schwartz 2022) Magar and Schwartz, Data Contamination: From Memorization to Exploitation, ACL 2022, https://aclanthology.org/2022.acl-short.18/. Test-set contamination separates into memorization (the model retains the contaminated data) and exploitation (it uses that data to score better downstream), with duplication count and model size affecting the two differently.
(Zheng 2023) Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena, NeurIPS 2023 Datasets and Benchmarks, arXiv:2306.05685, https://arxiv.org/abs/2306.05685. LLM judges exhibit position bias, verbosity bias, self-enhancement bias, and limited reasoning. A strong judge, once measured, can reach over 80% agreement with human preferences - about the level humans agree with each other.
(Panickssery 2024) Panickssery, Bowman and Feng, LLM Evaluators Recognize and Favor Their Own Generations, arXiv:2404.13076, https://arxiv.org/abs/2404.13076. Models score their own outputs higher than others' that humans rate equal, and the strength of that self-preference correlates linearly with the model's ability to recognize its own text.
(Shankar 2024) Shankar et al., Who Validates the Validators?, arXiv:2404.12272, https://arxiv.org/abs/2404.12272. LLM-generated evaluators inherit the problems of the models they evaluate and require human validation before their scores mean anything; also documents criteria drift, where the grading criteria users actually want depend on the outputs they have seen.
(Perez 2022) Perez et al., Red Teaming Language Models with Language Models, arXiv:2202.03286, https://arxiv.org/abs/2202.03286. Using language models to generate test cases against a 280B-parameter chatbot uncovered tens of thousands of offensive replies, private contact information, training-data leakage, and multi-turn harms. Explicitly one tool among several, not a complete safety answer.