Some more HCE "Texel Tuning" Data

[op12no2]

I'll note that this data proved terrible for NNUE. Since it was designed for HCE training specifically. Anything is better than nothing I suppose, but one can easily get much better data, with more diversity, on demand.

Copy that.

[AndrewGrant]

I'll note that this data proved terrible for NNUE. Since it was designed for HCE training specifically. Anything is better than nothing I suppose, but one can easily get much better data, with more diversity, on demand.

Copy that.

In fact, if you are using your own trainer of sorts, and would like to "sanity check" what you've done (indifferent from your eventual methods of generating your own data), hit me up on the OpenBench Discord and I will send the 100M fens that I shared with Koivisto team for the same purpose. I just ask that you would use it only to verify and see that part of the system is working -- not to produce then publish a set of weights.

[op12no2]

In fact, if you are using your own trainer of sorts, and would like to "sanity check" what you've done (indifferent from your eventual methods of generating your own data), hit me up on the OpenBench Discord and I will send the 100M fens that I shared with Koivisto team for the same purpose. I just ask that you would use it only to verify and see that part of the system is working -- not to produce then publish a set of weights.

OK thanks Andrew; early days here, but yes, all DIY Javascript as a learning exercise. I will eventually generate my own data but existing FEN datasets are a great way to get going quickly and play around with ideas.

[jdart]

So I have this naive implementation of the Texel tuning and it seems to work (just a local search). I tried implementing the Adam optimizer but the default learn rate is very small. It takes forever to tune values. It's like 1 centipawn per 100 epoch.

I use ADAM_ALPHA = 0.015 (initial learning rate). Works well for me.