跨国深谈

评审器 (judge)

13分钟阅读

跨国深谈← 返回←评审器 (judge)
评审

评审器 (judge)

概念
本站收录 3 集 · 14 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

14 条

我认为有很多它表现不太好的情况，但我觉得如果你把 LLM 当裁判的用途从一个基准重新定义为一个异常检测器，那它其实是可以的。
Well, I think there’s a lot of cases where it doesn’t work very well, but I think if you reframe the utility of an LLM as a judge from being a benchmark to being an anomaly detector, then I think it’s actually okay.
—— Ankur Goyal · [25:10]

指向原始笔记的链接

现在这听起来很吸引人，但它是一个非常危险的指标，因为很多时候，错误，它们只发生在长尾上，并且不经常发生，所以如果你只有 10% 的时间有错误，那么你可以很容易地通过让评判者一直说它通过来达到 90% 的一致性。
Now that sounds appealing, but it’s a very dangerous metric to use, because a lot of times, errors, they only happen on the long tail and they don’t happen as frequently, so if you only have the error 10% of the time, then you can easily have 90% agreement by just having a judge say it passes all the time.
—— Hamel Husain · [58:41]

指向原始笔记的链接

这是产品需求文档应该是什么样子的最纯粹的意义，就是这个评估判断器，它确切地告诉你它应该是什么，并且它是自动的且不断运行的。
This is the purest sense of what a product requirements document should be, is this eval judge that’s telling you exactly what it should be, and it’s automatic and running constantly.
—— Lenny · [61:35]

指向原始笔记的链接

我实际上使用 GPT-5.5 来给所有这些东西打分，因为它实际上是最严厉的评委。
I actually use GPT-5.5 to grade all these things because it’s actually the harshest judge.
—— 嘉宾 · [25:21]

指向原始笔记的链接

结果显示,我们的裁判至少与真正的专家临床医生不相上下,甚至可能略好一些。
The results showed that our judge is at least on par, if not slightly better, than the real expert clinicians.
—— Jared Joselowitz · [10:22]

指向原始笔记的链接

如果你只围绕某个特定的 LLM 裁判去爬山，你很可能是在对它过拟合。
And if you’re only hill climbing on a specific LLM judge, you’re probably overfitting on it.
—— Vishu · [07:35]

指向原始笔记的链接

你可以把它理解为一种方式：把你的技能转化成、或者说基于它们创建出非常聚焦、高准确率、非常快、非常便宜的「LM as judge」工具。
You can think of it as a way of turning your skills into or creating off of them very focused, high accuracy, very fast, very cheap LM as judge tools
—— Drew · [29:03]

指向原始笔记的链接

那是因为它们全都是针对测试去训练的。所以我们有一个理念：真正重要的是现实。现实是唯一一个你真正能信得过的裁判。
That’s because they all trained to the test. And so we had this philosophy that it’s really about reality. Reality is the only judge that you can actually trust.
—— Anastasios Angelopoulos · [03:35]

指向原始笔记的链接

瓶颈最终转移到了验证这一侧，因此，我们没有办法大规模地衡量价值，并以同样的速度判断质量。
the bottleneck ends up shifting to the verification side, and thus, we don’t have a way to measure value at scale and to judge quality at the same speed.
—— Maximillian Piras · [12:20]

指向原始笔记的链接

顺便说一句，这比大多数 LLM-as-a-judge 方法——也就是让一个 LLM 来判断那是优质的人类作品还是 AI 生成的 slop——表现都要好。
This performed better, by the way, than most LLM-as-a-judge methods of asking an LLM to judge if that is great human quality versus AI-generated slop.
—— Thais Castello Branco · [08:29]

指向原始笔记的链接

重要的一点是，验证者评判的是他们自己没写过的代码。
Important thing is the validators judge code that they didn’t write.
—— Tereza Tížková · [13:39]

指向原始笔记的链接

我觉得 GPT-3 在那个时候其实是相当有校准的。但因为是人类在评估模型有多好,它看起来非常好,因为人类是裁判,但我们一直在优化那个裁判,而不是自动化那一部分。
I think GPT-3 was actually quite calibrated back in that day. But because humans evaluate how good the models are, it looks really good because they are the judge, but we’ve been optimizing that judge instead of the automation part.
—— Diogo Almeida · [20:52]

指向原始笔记的链接

他们开始以某种方式利用测试数据，用 LLM 作为裁判来训练模型，用各种花招来刷基准，拿季度奖金之类的。
They start to kind of use leverage test data to use LLM as a judge for training of the models, use all the various tricks to basically gain benchmarks, get like quarterly bonuses and so on.
—— Alex Mashrabov · [23:52]

指向原始笔记的链接

所以我认为历史会回望那些人，如果不让它发生，会给他们非常差劲的评价。
So I think history will look back on people and judge them very poorly if they don’t let it happen.
—— David George · [28:22]

指向原始笔记的链接

② 出现在这些集

3 集

③ 关联

点进去有真内容 —— 本页主要出口

Portola · Tolan · LLM · 提示词 · 记忆 · 响应时间 · 即兴演员 · Quintin · Jared Joselowitz · 设定种子

反向链接