Michael Pang

It looks like the model might be overfitting to the opening before it has a chance to learn basic chess principles like piece values. GM games might not be the...

Yes, sort of like AlphaZero :) That way we can also see where it plateaus and iterate faster. BTW can you comment on why you closed my PR? I can...

Are you referring to the "data format" issue I raised? I've already fixed all the stuff I mentioned (and a few other bugs) so after a bit more testing I...

Noted thanks! I hope we can share test cases too. The training loss is questionable (in my opinion) because it can just overfit the common openings, and we don't know...

Sure I'll look around. Also could you see if you can use your company can let you use their Azure server for FTP? BTW if you change the keras.json to...

~~It looks like it's private?~~ https://github.com/crypt3lx2k/Zerofish

You have to go up at least one directory, so the chess_zero folder is in the working directory of run.py.

We could just post all our training data publicly so anyone is free to experiment with different architectures on their hardware

Comparing the two sampling implementations: https://github.com/minimaxir/gpt-2-simple/blob/ca6bc61d958fd4c474af9a412ace27279b88dd90/gpt_2_simple/src/sample.py#L41 https://github.com/AIDungeon/AIDungeon/blob/e87d13b6b011a88c2fb034617ffef2ee4fef98e3/generator/gpt2/src/sample.py#L56 There are some minor differences, like how initial loop variables are computed, but for the most part they're the same. Another source of...

Lower temperature should lead to more coherent (but more repetitive) output. I believe they used a DGX-1 to finetune 1.5B