https://dailypypy.org/tags/evaluation/ Recent content in Evaluation on 每日拍拍 Hugo -- gohugo.io zh-tw © 2026 每日拍拍 Mon, 19 Oct 2026 12:30:00 +0800 155076163427069952 154825760438254592 https://dailypypy.org/learn/mlx-lm-evaluation-regression/ Mon, 19 Oct 2026 12:30:00 +0800 https://dailypypy.org/learn/mlx-lm-evaluation-regression/ <!--- 1440x768 prompt: masterpiece, best quality, highres, clean anime illustration, japanese anime style, soft shading, flat color design, 1girl, black hair, green eyes, white off-shoulder shirt, black short skirt, eye-level frontal waist-up portrait, holding a closed unmarked notebook against her chest with both hands, playful winking smile, looking at viewer, pastel pale rose and soft seafoam background, neat composition, detailed eyes, cute and smart vibe, minimal background, polished illustration, no text negative prompt: worst quality, bad eye, bad hand, extra limbs, manga, multiple views, monochrome, text, signature dedup note: Distinct regression-governance follow-up to ../mlx-lm-batch-inference/ and ../mlx-lm-quantization-convert/. Those cover generic runners, sampling, quantization/speed/quality comparisons. This article instead centers immutable case identity, suite and scoring-protocol hashes, independently rescored raw artifacts, fail-closed paired transitions, critical-case gates, failure grouping and reviewed baseline promotion. It does not repeat ../mlx-lm-structured-json-validation/ constrained decoding, Pydantic or repair/retry. Strict 195-post current-source inventory checked before drafting. source check: MLX-LM official README and sample_utils.py reviewed 2026-10-19. Gate policies and synthetic fixtures below are original example application contracts, not official MLX-LM benchmark recommendations. No model-quality or throughput measurements are claimed. verification: Python 3.14.3; all five Python blocks parsed and 27 pytest cases passed, including article test, malformed materials, paired transitions, exit codes and the actual capture adapter under a synthetic engine. No real-model quality claim. Current-source max character 5-gram Jaccard 0.06270 against mlx-lm-batch-inference; manual scope review confirms distinct regression-governance task. Cover prompt cc9e9146-8c1f-4a9a-9c87-98068e42d1ba produced DAILYPYPY_00140_.png, verified PNG 1440x768 and visually inspected for text, face, hand and artifact issues. publication note: Required random display date retained; publishDate permits this morning release with buildFuture=false. ---> <h2 class="relative group">一. 前言：總分沒掉，怎麼還是壞了？ <div id="一-前言總分沒掉怎麼還是壞了" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 ltr:-left-6 rtl:-right-6 not-prose group-hover:opacity-100"> <a class="group-hover:text-primary-300 dark:group-hover:text-neutral-700" style="text-decoration-line: none !important;" href="#%e4%b8%80-%e5%89%8d%e8%a8%80%e7%b8%bd%e5%88%86%e6%b2%92%e6%8e%89%e6%80%8e%e9%ba%bc%e9%82%84%e6%98%af%e5%a3%9e%e4%ba%86" aria-label="定位點">#</a> </span> </h2> <p>拍拍君替本機模型換了一版 Prompt，測試分數還是 75%。 看起來可以放心？先等等。 舊版答對的關鍵題被答錯，新版只是救回另一道普通題。 <strong>總分相同，不代表沒有回歸。</strong></p>