Stockfish 14 release round the corner

[dkappe]

Just to correct you: Fat Fritz 2 was not trained on lc0 training data.

I don't think you can assert that. And I would sooner assert the opposite, given Albert's inability to do anything on his own.

I can and I did. Best to sit down now and not speak of things about which you know nothing.

As per Albert himself, it was trained on Fat Fritz 1 data. Fat Fritz 1 is just rebranded Lc0. According to Albert, the network was pretrained on human game results, but the majority of the training data did result from Lc0 self play with some tweaked hyperparameters from what I can discern from ChessBase's articles on it and Albert's comments.

What do you mean by “Lc0 self play?”

The Lc0 engine playing against itself (using all the Lc0 team's self play game data generation code)

Try to remember the context of my initial post. If the Stockfish team generated their own data by way of a separate training run (maybe with some SL pretraining on some number of SF games), I don't think that would have any meaningful impact on the final NNUE network or whether it was deemed kosher by CCRL.

OK. Then we are clear. This self play data is what I refer to as lc0 training data. I have found this sort of data to be less than satisfactory, probably because of the use of temperature. Data generated through a UCI harness (or in the case of Night Nurse, through a julia/torchscript engine using Bad Gyal) where the opening randomization is taken care of apart from the engine, yields better results. See my post in the leela data thread.

[Sopel]

There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well.

But that's basically what stockfish 14 used for training...

[dkappe]

Since you appear not to be that familiar with lc0, I’ll cut you some slack. There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well. If you instead use some sort of UCI harness that takes care of randomized openings, you can generate nnue training data that works much better. I hope you now understand the distinction.

Looks like I caught the word games you put down. I encourage everyone to read and re-read postings from Dkappe and Albert. These sort of shenanigans are just some of the tools they use to peddle their views while having a degree of truth to fall back onto. You see it time and time again on the forums, and in the world at large.

Are you having a stroke? It’s really not that hard to understand.

[dkappe]

There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well.

But that's basically what stockfish 14 used for training...

Yes. My point is that that training data isn’t as good as data generated via some sort of UCI harness using no temperature, randomized openings and direct use of EGTB rather than rescoring.

[Sopel]

Since you appear not to be that familiar with lc0, I’ll cut you some slack. There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well. If you instead use some sort of UCI harness that takes care of randomized openings, you can generate nnue training data that works much better. I hope you now understand the distinction.

Looks like I caught the word games you put down. I encourage everyone to read and re-read postings from Dkappe and Albert. These sort of shenanigans are just some of the tools they use to peddle their views while having a degree of truth to fall back onto. You see it time and time again on the forums, and in the world at large.

Are you having a stroke? It’s really not that hard to understand.

Let me get this straight. You're claiming that data generated with Lc0 used as an UCI engine is not Lc0 data?

[dkappe]

Since you appear not to be that familiar with lc0, I’ll cut you some slack. There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well. If you instead use some sort of UCI harness that takes care of randomized openings, you can generate nnue training data that works much better. I hope you now understand the distinction.

Looks like I caught the word games you put down. I encourage everyone to read and re-read postings from Dkappe and Albert. These sort of shenanigans are just some of the tools they use to peddle their views while having a degree of truth to fall back onto. You see it time and time again on the forums, and in the world at large.

Are you having a stroke? It’s really not that hard to understand.

Let me get this straight. You're claiming that data generated with Lc0 used as an UCI engine is not Lc0 data?

I am claiming two things:

1. NNUE training data generated via a UCI harness is not lc0 training data (the sort you would download from the lczero web site). You couldn’t use it in the lc0 training pipeline.
2. Data generated with a specific net should really be called by the name of that net. For example, data generated using bad gyal 9XL using a0lite julia is not a0lite julia data but rather bad gyal 9XL data.

[Sopel]

Since you appear not to be that familiar with lc0, I’ll cut you some slack. There is a special mode in lc0 for generating training data via self play. This type of data when converted to nnue training data doesn’t work that well. If you instead use some sort of UCI harness that takes care of randomized openings, you can generate nnue training data that works much better. I hope you now understand the distinction.

Looks like I caught the word games you put down. I encourage everyone to read and re-read postings from Dkappe and Albert. These sort of shenanigans are just some of the tools they use to peddle their views while having a degree of truth to fall back onto. You see it time and time again on the forums, and in the world at large.

Are you having a stroke? It’s really not that hard to understand.

Let me get this straight. You're claiming that data generated with Lc0 used as an UCI engine is not Lc0 data?

I am claiming two things:

1. NNUE training data generated via a UCI harness is not lc0 training data (the sort you would download from the lczero web site). You couldn’t use it in the lc0 training pipeline.
2. Data generated with a specific net should really be called by the name of that net. For example, data generated using bad gyal 9XL using a0lite julia is not a0lite julia data but rather bad gyal 9XL data.

1. So basically yes, you are claiming what I said.
2. That's how money laundering works, yes. It also means that stockfish was not trained on lc0 data.

[dkappe]

I am claiming two things:

1. NNUE training data generated via a UCI harness is not lc0 training data (the sort you would download from the lczero web site). You couldn’t use it in the lc0 training pipeline.
2. Data generated with a specific net should really be called by the name of that net. For example, data generated using bad gyal 9XL using a0lite julia is not a0lite julia data but rather bad gyal 9XL data.

1. So basically yes, you are claiming what I said.
2. That's how money laundering works, yes.

I don’t understand your response to #2.

As for #1, you are about to miss an opportunity to improve the SF net because you dislike me and aren’t open to what I’m saying. I’ve been training NNUE nets from mcts/nn data longer than anyone. I’m telling you that if you do direct generation of NNUE training data via a UCI harness instead of converting lc0 training data, you will have much better results. You might be able to use the self play training function without temperature and a very large opening book and direct use of EGTB instead of rescoring. I haven’t looked at lc0 in a while, so don’t know what the opening book and egtb support looks like for self play data gen.

I don’t expect much from Andy, but you have always struck me as a fairly rational fellow. But if you want to continue this semantic food fight, do it without me.

[Madeleine Birchfield]

The interesting thing about the current argument between the usual suspects is that if either side is correct, then CCRL and other rating lists would have to test Stockfish 14. If dkappe is right and Stockfish did not use Leela data, then there would be no problem with CCRL testing Stockfish 14. If dkappe is wrong and Stockfish did use Leela data, then by previous precedent, such as testing Allie with a Leela net, testing Nemorino 6.00 which used Stockfish data, testing BBC 1.4 which used a Stockfish net, testing Fat Fritz 2 which used Leela data, and so forth, then there would be no problem with CCRL testing Stockfish 14.

The alternative of course would be to remove Fat Fritz 2, Allie, and Nemorino 6.00 from the ratings list, but they are loath to remove any of them.

[Sopel]

I am claiming two things:

1. NNUE training data generated via a UCI harness is not lc0 training data (the sort you would download from the lczero web site). You couldn’t use it in the lc0 training pipeline.
2. Data generated with a specific net should really be called by the name of that net. For example, data generated using bad gyal 9XL using a0lite julia is not a0lite julia data but rather bad gyal 9XL data.

1. So basically yes, you are claiming what I said.
2. That's how money laundering works, yes.

I don’t understand your response to #2.

As for #1, you are about to miss an opportunity to improve the SF net because you dislike me and aren’t open to what I’m saying. I’ve been training NNUE nets from mcts/nn data longer than anyone. I’m telling you that if you do direct generation of NNUE training data via a UCI harness instead of converting lc0 training data, you will have much better results. You might be able to use the self play training function without temperature and a very large opening book and direct use of EGTB instead of rescoring. I haven’t looked at lc0 in a while, so don’t know what the opening book and egtb support looks like for self play data gen.

I don’t expect much from Andy, but you have always struck me as a fairly rational fellow. But if you want to continue this semantic food fight, do it without me.

By 2. I mean that you described an easy way to make every engine's net original, by your standards.

Also I have nothing against what you're claiming regarding to what works with NNUE training or doesn't. However I prefer following results rather than claims in my research, especially if it involves large resource requirements.