Leela data worth +20 elo

[dkappe]

Looks like a test by the SF team has shown that leela + sf data is 20 elo stronger than sf data by itself. +20 elo is no small thing at this level.

[Werewolf]

Source?

I was wondering if SF 14.2 or SF 15 was due soon based on the time gap between 14 and 14.1

[dkappe]

Source?

I was wondering if SF 14.2 or SF 15 was due soon based on the time gap between 14 and 14.1

https://tests.stockfishchess.org/tests/ ... 51c74890b8

More like 16.5 elo.

Whenever they get +30 elo, they’ll release. And if we drop Dragon 3, they drop SF 15.1.

[Uri Blass]

Source?

I was wondering if SF 14.2 or SF 15 was due soon based on the time gap between 14 and 14.1

https://tests.stockfishchess.org/tests/ ... 51c74890b8

More like 16.5 elo.

Whenever they get +30 elo, they’ll release. And if we drop Dragon 3, they drop SF 15.1.

I do not understand what is considered to be Leela data.
I assume both versions use NNUE.

Does one of them use NNUE that is based on Leela's code when another one is using NNUE that is not based on Leela's code.

[Vernon Crawford]

The CCRL's most recent policy has been to exclude valid engines because their NNUE was not created from its own engines games.
Interestingly enough, even though SF and Leela share data they are eagerly tested by CCRL.

what hyprocrisy
my guess: a new convoluted and ignorant policy will soon be communicated which attempts to justify their decisions in this regard..
after which connor mcmonicle with chime in to defend the illogical with some obtuse tech jargon mumbo jumbo.

[AndrewGrant]

The CCRL's most recent policy has been to exclude valid engines because their NNUE was not created from its own engines games.
Interestingly enough, even though SF and Leela share data they are eagerly tested by CCRL.

what hyprocrisy
my guess: a new convoluted and ignorant policy will soon be communicated which attempts to justify their decisions in this regard..
after which connor mcmonicle with chime in to defend the illogical with some obtuse tech jargon mumbo jumbo.

I'm pretty confident that CCRL does not have that policy.

[connor_mcmonigle]

The CCRL's most recent policy has been to exclude valid engines because their NNUE was not created from its own engines games.
Interestingly enough, even though SF and Leela share data they are eagerly tested by CCRL.

what hyprocrisy
my guess: a new convoluted and ignorant policy will soon be communicated which attempts to justify their decisions in this regard..
...

CCRL has no such official policy in my understanding, though CCRL testers have stated that they're generally uninterested in testing engines which copy Stockfish's NNUE implementation+training code and use weights produced by way of training on data generated by other engines. CCRL already has tested engines trained on data generated by other engines such as Expositor.

Regardless, it shouldn't exactly come as a surprise that testers would have some interest in testing the definitively strongest engine regardless of the extent to which it is "original" (which Stockfish certainly is for any reasonable definition of "original").

...
after which connor mcmonicle with chime in to defend the illogical with some obtuse tech jargon mumbo jumbo.

I'm happy to answer any clarifying questions you might have if you're struggling to understand something I've written...

[Vernon Crawford]

CCRL has no such official policy in my understanding...

If engine B copies engine A's evaluation function functionally verbatim, can we really consider engine B a unique engine? At this point, why not just copy the search as well? The majority of authors have avoided doing this as it's generally regarded as incredibly boring and doing so eliminates any room for creativity and differentiation with respect to the evaluation function.

Current exceptions that I'm aware of are: FireNN, Gogobello, Mayhem, Tornado, AndscacsNNUE (the author mentioned this was just an experiment and not to be tested), BBC (also just an experiment), Caissa and Coiled. It could make sense to append "SF" to the end of these engines' names, but I'd argue you'd be better off just not testing them in that case (which is the policy CCRL has adopted FWIW).

So you deny they have such a policy on one hand, but specify they actually do on the other hand?
Can people actually believe anything you say?

prepare: techno mumbo jumbo confusing smoke-screen!

[dkappe]

What is “leela data?” Leela generates huge numbers of self play games for training. This has score, result, position, move and stuff like moves left and policy. The leela project wrote a tool for the SF project that converted leela data to data that is useful for training NNUE. I think over 50 billion positions of leela data have been converted.

So the leela data is huge and of very high quality. You need lots of positions, 10 billion plus, to train these big nets. More like 20 billion if you want to be safe.

[Vernon Crawford]

The CCRL's most recent policy has been to exclude valid engines because their NNUE was not created from its own engines games.
Interestingly enough, even though SF and Leela share data they are eagerly tested by CCRL.

what hyprocrisy
my guess: a new convoluted and ignorant policy will soon be communicated which attempts to justify their decisions in this regard..
after which connor mcmonicle with chime in to defend the illogical with some obtuse tech jargon mumbo jumbo.

I'm pretty confident that CCRL does not have that policy.

Not surprised you jumped in here, along with Connor to defend CCRL...seeing how they seem to have become your lapdog, due to criticisms, not to mention boycott threats.

Do I need post them here?