Let the AI pick a translation better than the "right answer

<p><strong>Instead of imitating one sacred right answer, have it produce N candidates and pick the best — but it's no cure-all</strong></p> <p><strong>There is one right answer, and it is sacred</strong> — that article of faith in AI training cracked. Have the teacher produce N candidates and let a quality estimator pick among them, and what you get is a set of model answers that beats the teacher's own first shot (+2 on Live; 90→96 together with the CPO that followed).</p> <p>But the thing I'm sure of is the flip side. <strong>If the yardstick you select with is off target, you lose outright to a single clean answer.</strong> When I selected by closeness to a right answer that faced the wrong direction, 82% fell to 73–77%.</p> <p>More volume doesn't change that. Data inflated to 8,553 examples by raising the temperature lost to a clean 3,301. <strong>The question to ask wasn't "what should N be" — it was "does this yardstick point at the kind of good I actually want?"</strong></p> <p>When you teach an AI something, the most straightforward approach is this: <strong>prepare one correct set of model answers and have it imitate them</strong>. For translation, you show it a great many pairs that say "the right answer for this sentence is this translation." Straight out of the machine learning textbook, and it's how I worked for a long time.</p> <p>Behind it sits a quiet article of faith: the assumption that <strong>"there is one right answer (gold), and it is sacred."</strong> Raise the quality of the model answers and you raise the quality of the student. So you carefully prepare one good right answer —</p> <p>Then one experiment cracked that faith. <strong>Having the model produce N candidates and picking the "best one" gave better translations than having it imitate a single right answer.</strong> And another experiment from the same period taught me, just as clearly, that this way of picking is <strong>no cure-all</strong> — get the yardstick wrong and you lose outright to one clean set of model answers. This article is about both sides.</p> <h2> Where it started — one weak mode I just couldn't close </h2> <p>LYR's small model has three places it gets used (modes): <strong>Live</strong>, which translates video subtitles in real time; <strong>Page</strong>, which translates the whole screen; and <strong>Manga</strong>, which translates comics.</p> <p>Of the three, <strong>Live</strong> was stubbornly the weak one (what I'm training here is the mid-size one among the small models — the <strong>8B specialist</strong>). Same small model, and yet Page and Manga were roughly level with the large model above them, while Live alone was left well behind. Looking at the causes one case at a time, the shape of the failures came into view: idioms translated literally (reading "isn't like you" as the opposite of 「あなたらしくない」, the idiom it should have produced), the wrong subject picked, words inserted that were never there.</p> <p>The first move that comes to mind is "collect more gold data for Live and have it imitate more of it." But who writes that gold data? <strong>The larger model above it — the teacher — writes it</strong> (the distillation from <a href="https://lyr.jp/en/articles/specialist-02-fine-tuning/" rel="noopener noreferrer">last time</a>). You have the teacher translate once and use that as the model answer. Except the teacher's own translations sometimes missed the idiom. <strong>If the model answers miss it, the student misses it too.</strong></p> <h2> The realization — the teacher's "first shot" isn't necessarily its best </h2> <p>So I changed the framing. Stop having the teacher translate <strong>exactly once</strong> and then worshipping that as the one right answer. Instead —</p> <p><strong>Have the teacher translate the same sentence N different ways. Then mechanically pick the "best translation" out of that set.</strong></p> <p>That's the idea behind <strong>best-of-N</strong> (take the