Stockfish network tuning

[jdart]

Not being deeply plugged into the project, I'm having a little hard time understanding what Stockfish has been doing for network tuning these days.

I hear it is now using Lc0 data - I guess this is Lc0 self-play? t seems the Lc0 project is generating copious data for its own tuning needs. How is this used by SF? Does it tune on game result, or eval or both? What size data set is it tuned on? (Anyone can do their own network tuning, but I guess I am asking about how commonly network updates in the official branch are made).

[dangi12012]

Best you ask them on their discord!
They are really nice and have a special questions channel.

[connor_mcmonigle]

https://tests.stockfishchess.org/nns
Here's a list of all networks uploaded to fishtest with networks passing SPRT highlighted in green which should give some idea of the frequency at which the default network is updated. Alternatively, you can check the commit history to see when the default network is updated. This might be misleading though as many of the networks tested correspond to SPSA tunes on top of existing networks.

[Sopel]

I hear it is now using Lc0 data - I guess this is Lc0 self-play?

yes

t seems the Lc0 project is generating copious data for its own tuning needs. How is this used by SF? Does it tune on game result, or eval or both?

Recently pretty much purely on eval, since we started skipping positions based on how close the eval is to the result (the less likely the result is from the current eval the higher chance to skip a position). Before that the best was around 0.8 * eval + .2 * result.

what size data set is it tuned on?

Around 10B-100B positions. Depending on how many datasets you want to throw at it. You can get good results with around 10-20B.

(Anyone can do their own network tuning, but I guess I am asking about how commonly network updates in the official branch are made).

Very irregularly.

[dangi12012]

I hear it is now using Lc0 data - I guess this is Lc0 self-play?

yes

t seems the Lc0 project is generating copious data for its own tuning needs. How is this used by SF? Does it tune on game result, or eval or both?

Recently pretty much purely on eval, since we started skipping positions based on how close the eval is to the result (the less likely the result is from the current eval the higher chance to skip a position). Before that the best was around 0.8 * eval + .2 * result.

what size data set is it tuned on?

Around 10B-100B positions. Depending on how many datasets you want to throw at it. You can get good results with around 10-20B.

(Anyone can do their own network tuning, but I guess I am asking about how commonly network updates in the official branch are made).

Very irregularly.

Where is the source for your wild claims?
This guy sopel does make up stuff 99% of the time so - take that with a big chunk of salt since he is a known forum troll here..
Without URLS to the actual source you shouldnt believe sopel.

[Guenther]

I hear it is now using Lc0 data - I guess this is Lc0 self-play?

yes

t seems the Lc0 project is generating copious data for its own tuning needs. How is this used by SF? Does it tune on game result, or eval or both?

Recently pretty much purely on eval, since we started skipping positions based on how close the eval is to the result (the less likely the result is from the current eval the higher chance to skip a position). Before that the best was around 0.8 * eval + .2 * result.

what size data set is it tuned on?

Around 10B-100B positions. Depending on how many datasets you want to throw at it. You can get good results with around 10-20B.

(Anyone can do their own network tuning, but I guess I am asking about how commonly network updates in the official branch are made).

Very irregularly.

Where is the source for your wild claims?
This guy sopel does make up stuff 99% of the time so - take that with a big chunk of salt since he is a known forum troll here..
Without URLS to the actual source you shouldnt believe sopel.

The one forum troll I see here is always you the giganut! Stay out of threads you have no clue about and don't stalk sopel anymore.
You are endlessly pestering the programmers section... ignored by most of all here already.

[ChickenLogic]

I hear it is now using Lc0 data - I guess this is Lc0 self-play?

yes

t seems the Lc0 project is generating copious data for its own tuning needs. How is this used by SF? Does it tune on game result, or eval or both?

Recently pretty much purely on eval, since we started skipping positions based on how close the eval is to the result (the less likely the result is from the current eval the higher chance to skip a position). Before that the best was around 0.8 * eval + .2 * result.

what size data set is it tuned on?

Around 10B-100B positions. Depending on how many datasets you want to throw at it. You can get good results with around 10-20B.

(Anyone can do their own network tuning, but I guess I am asking about how commonly network updates in the official branch are made).

Very irregularly.

Where is the source for your wild claims?
This guy sopel does make up stuff 99% of the time so - take that with a big chunk of salt since he is a known forum troll here..
Without URLS to the actual source you shouldnt believe sopel.

Yeah better not believe the guy who was responsible for every change of nnue architecture in Stockfish (after the initial halfkp) and documents all the experiments he does in detail. Just about 500 pages long: https://docs.google.com/document/d/1gTl ... 6jvNqqMkAY

I seriously don't get why you're still allowed to post in this forum. Willfully spreading misinformation just because you dislike Sopel. How about not being an a**hole?

[dangi12012]

Look all I did was ask for a source for his wild random numbers he quoted. He did not provide them so for. Sopel likes to hate on multiple other people on hundreds of other threads. This is very discouraging for people who are new that the first post that comes is from a troll like sopel who likes to hate on things.

For example he is still claiming that opening books are totally useless:

Who in their right mind even needs to query ALL the games for a position AT ONCE? This is trying to optimize something that no one will ever use, while completely removing the possibility of partial queries.

While this is literally the core of chess databases...

No one should take anything he says seriously. Look at his posting history. 99% negative hating on others. He is making claims. He should back them up. Its really that simple.

[ChickenLogic]

Look all I did was ask for a source for his wild random numbers he quoted. He did not provide them so for. Sopel likes to hate on multiple other people on hundreds of other threads. This is very discouraging for people who are new that the first post that comes is from a troll like sopel who likes to hate on things.

For example he is still claiming that opening books are totally useless:

Who in their right mind even needs to query ALL the games for a position AT ONCE? This is trying to optimize something that no one will ever use, while completely removing the possibility of partial queries.

While this is literally the core of chess databases...

No one should take anything he says seriously. Look at his posting history. 99% negative hating on others. He is making claims. He should back them up. Its really that simple.

Please back up your wild claim that 99% of his posts are just hating on other people. So far I didn't see any back up of this claim. You should back up your claims or you're just a troll. Thanks and good luck reading all the messages he ever wrote. It should be really simple.

[Rebel]

Look all I did was ask for a source for his wild random numbers he quoted.

You did more than that, you out of the blue on a neutral post insulted him.

He did not provide them so for.

No need, known by insiders.

Sopel likes to hate on multiple other people on hundreds of other threads. This is very discouraging for people who are new that the first post that comes is from a troll like sopel who likes to hate on things.

Sopel, Sopel, Sopel, go away man, find yourself another forum.