Actually to me it sounds it could be benchmarked if this kind of effect exists in the first place.
To completely squash the issue, a few cheap LoRa iterations will do the trick just fine.
I think we can all agree that Jev is not rocket science. It's a good idea executed well, with marketing that might have been a tad too bold
And further down "TypeSafe computes confidence from how the probability is spread across the options. All of it on one option gives 1.0; the more evenly it spreads, the lower the confidence. This demo uses (3 × largest probability − 1) / 2 to approximate confidence for three options."
So while we don't know the exact formula they use, it is just a function over the probabilities
I am open to the argument that this does not work well if you just plug in a qwen model instead of a model that is trained to output more statistically useful token distributions
we agree then, that is the entirety of my argument. Getting a deep net especially one that is anywhere near even SLM size to be calibrated is tough, especially across domains. They claim calibration across a variety of datasets which is interesting.
It means confidence is just a converted max probability and not an independent signal.
I ran some tests using GPT-4 to do some basic classification a couple years ago. On ambiguous options which had to be escalated to a human, the LLM would regularly output something like a 99.8% probability, compared to 99.99% for a correct answer.
https://til.simonwillison.net/llms/llama-cpp-python-grammars
I got this technique to work extremely reliably last year. However there were a bunch of caveats: 1) Firstly, you must institute a check that the multiple choice tokens dominate the output distribution. They should sum to 95% or more, ideally 99%, or the LLM is not following instructions properly. This is also the problem with constrained decoding - if the LLM really doesn't want to output a valid answer, the one you extract will not be high quality. 2) You need to ask it multiple times, permuting which option corresponds to which letter, and average the results. LLMs are surprisingly biased towards picking "A", especially if they're otherwise not sure. 3) For the same reason, performance improves if you frame the prompt as if it were the middle of a quiz. "Question 1" carries baggage that "Question 12" doesn't. 4) You must be exceedingly careful with tokenization.
But when all was said and done, I got a general purpose A/B classifier that gave high resolution quantitative output for the cost of a couple dozen tokens ingested and a couple inference passes.
The whole point of my argument is that neither is good, but from a technical perspective logprobs is probably the worst unless you train a model on specific outputs. In which case you'd throw out the generality again, so when I think about it more, it's actually the worst overall. In my experiments, having the model simply assign "high" or "low" probability in a structured output generally performs best. You can try numbers, but you will never get anything close to what you could expect from traditional ML. And most certainly not from logprobs.
GP pointed at a causal explanation for this: almost every sentence in English that's a statement will start with "A" or "An", so "biased towards picking ''A''" will include most attempts at saying anything long-form for any reason.
Meanwhile, the bias could be as much as 70% in favor of A in ambiguous cases - a signal completely drowning the <1% inclination to violate the format.
Not nearly as sophisticated as myself who would mutter "When in doubt - Charlie out" before marking C.
Also, Jev/laya do it in one forward pass, for multiple questions about the same state, rather than multiple passes for one question about that state. Well, for the usual multilingual configuration, two forward passes through different small models for laya, but that's because one is the router which chooses which model should do the real work, but still.
I contribute my experience here only because I've seen a lot of chatter lately about doing exactly this sort of thing, and I thought I'd share how I made it work for me. There are a lot of ways it can silently fail and give bad numbers if you aren't careful, and I wouldn't want people to think it doesn't work just because they used a vibe coded GitHub project from the last 48 hours that doesn't take these things into account.
Prompt part: "What is better, toast or bread?"
Incomplete answer part: "The answer to this question is "
and then have the LLM finish the answer. I did this with subtitle translation using llama.cpp (with Python) and had great success. Just past 5 already translated subtitles as the incomplete answer, and the LLM infallibly just continues to translate. No markdown, and usually no talkback if the subtitles contain nasty subjects like bioweapons or nuclear stuff. It just works.
There, I fixed your problem.