Hacker News Reader

Comment by ramon156 | Hacker News Reader

Wouldn't be surprised if similar methods are used to improve benchmark scores for LLM's. Just make the LLM respond correctly on popular questions

View on HN

Oh for sure. The questions for most benchmarks can be downloaded on hugging face.

Plus rankings like lmsys use a known fixed system prompt

I thought this is why most of the benchmarks have two parts, with one set of tests public and the other set private?

In an ideal world yes, but in order for LLM authors to provide the evals they need to access the private set (and promise not to train on them or use that to influence eval/training methods).

Either the eval maintainers need to be given the closed source models (which will likely never happen) or the model authors need to be given the private evals to run themselves.

So the entire benchmark scheme is worthless?

Well it depends on how you define worthless. For you as an individual to ascertain truth, it may be useless. To build up a bloated AI Enterprise stock value. For false consensus narrative scripting. Very valuable.

It’s essentially worthless for you, as a consumer of them. The best way to see which one works best is to give a bunch of them a try for your specific use case

Not if you have a private eval you use for your own use cases.

Benchmarks have been close to worthless since Nvidia cheated 3DMark benchmarks by adding benchmark detecting code to their drivers in the early 2000s [1].

[1] https://www.eurogamer.net/news280503nvidia

That's a bit dramatic. One person cheating doesn't mean everybody is cheating. Cycling didn't become pointless after Lance Armstrong was caught doping. But it may be the tip of the iceberg and warrant a change in methodology or further investigation.

Were there like 20 or 30 other cycling cheaters that were caught alongside lance Armstrong?

Alberto Contador (2010, ban upheld in 2012) – Stripped of his 2010 Tour de France title due to clenbuterol use. Jan Ullrich (2012) – Officially found guilty of doping in connection with Operation Puerto, though his offenses dated back to the 2000s. Frank Schleck (2012) – Tested positive for a banned diuretic during the Tour de France. Johan Bruyneel (2018) – Armstrong’s former team director was banned for life for his role in systematic doping. Chris Froome (2017 case, cleared in 2018) – Found with high levels of salbutamol; later cleared by the UCI and WADA. Jarlinson Pantano (2019) – Tested positive for EPO and received a four-year ban. Nairo Quintana (2022) – Disqualified from the Tour de France for tramadol use, though it was not classified as a doping offense.

The sport was tainted before Lance and still is.

NVIDIA isn’t the only one who’s been caught.

https://www.anandtech.com/show/15703/mobile-benchmark-cheati...

> Either the eval maintainers need to be given the closed source models (which will likely never happen)

Given that the models are released to public, the test maintainers can just run the private tests after release, either via the prompts or via an api. Cheating won't be easy.

Closed source in this context literally means that they are not released.

They dont need them to be released(in the sense that you have a copy of the binary) to evaluate the model. The costly model training is useless unless access is given to people who pay for it.

The models of Open AI, Claude and other major companies - are all available either for free or a small amount(200$ for OpenAI Pro). Anyone who can pay this, can run private tests and compare scores. So, the public does not need to rely on benchmark claims of OpenAI based on its pre-release arrangements with test companies.

> Anyone who can pay this, can run private tests and compare scores.

Yes, by uploading the tests to a server controlled by OpenAI/Anthropic/etc

These prompts are fielding millions of queries. The test questions are a small part of them. Further, the server doesn't know if it got the right answer or not, so it can't even train on them. Whereas in the arrangement with the testing companies before release, they can potentially do so, as the they are given the scores.

Because of how LLMs generalize I'm personally of the opinion we shouldn't have public sets anymore.

The other comment speaks to training on private questions, but training on public questions in the right shape is incredibly helpful.

Once upon a time models couldn't produce scorable answers without finetuning on the correct shape of the questions, but those days are over.

We should have completely private benchmarks that use common sense answer formats that any near-SOTA model can produce.

Looking forward to LLMgate

Reply on news.ycombinator.com