In the Leela Chess project, we generate a huge amount of data. We use them to generate the network files to use with Lc0 for further data generation, but also with other chess engines, like Ceres. The same data are often used by individual project contributors to generate additional network files using the “supervised learning” approach.
Our intention has always been for “our” data to be open and available to everyone to use. To that end, we adopted an open license to allow their wide use:
Therefore we are very pleased that Stockfish, starting from today, is using a NNUE network file trained on the same data
Both projects have mentioned before that our “teams will join forces to demonstrate our commitment to open source chess engines and training tools, and open data.” This is the first concrete result stemming from this effort, and we promise it won’t be the last.
This might be of great use for training networks and tuning evaluations.
It always have been. Efforts were made to keeping those data available for everyone, the only problem is few people don't seem to willingly pay any tribute or even care enough to spell people's last name right.
So they already did the equivalent of what AS did for FF2? Looks already stronger and it can only get better. Time for a SF14 that will retake ALL the first spots in the rating lists.
So I started training Night Nurse from data generated by Bad Gyal 8 (then 9) as an exercise to see what kind of nnue net a mcts/nn engine would spawn. This data was generated using uci over a set of random openings and also a very large set of human < +-200 cp openings.
After generating a large amount of this data, I thought about converting all the “free” Bad Gyal self-play training data I had sitting about. A few lines of python later and I had maybe 250m positions I could add to my existing 300m positions. Instant elo boost, right?
Nope. They added maybe 10 elo. Testing with just the training data, it was maybe 80 elo weaker than a nnue net trained on the non-training data.
Now I had found a sweet spot of lambda = 0.7 for the Bad Gyal data. Moving to lambda 1.0 reduced the difference to 20 elo, but didn’t wipe it out, and the resulting nets were weaker than the 0.7 nets. My hypothesis is that the use of temperature makes the data perform worse when used with lambda < 1.0.
There’s quite a bit to critique about my experiment — data that doesn’t exactly match the source nets, etc. — but the difference was big enough that I stopped using training data as a source.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
