search

LeelaChessZero / lc0

The rewritten engine, originally for tensorflow. Now all other backends have been ported here.
GNU General Public License v3.0
2.37k stars 523 forks source link

Lc0 and fine tuning #1946

Closed StepHaze closed 5 months ago

StepHaze commented 7 months ago

Hello, is it possible what I would like to do? I'd like to take a Lc0's pre-trained model and train it on new data (1000 chess games of mine), so it will play in my style and my strength?

Thanks!

brianprichardson commented 7 months ago

It is not possible as a practical matter. Several hundred thousand games would be required for a net to begin to learn a particular style.

loneicewolf commented 6 months ago

it was a interesting question though, @StepHaze

StepHaze commented 6 months ago

It is not possible as a practical matter. Several hundred thousand games would be required for a net to begin to learn a particular style.

Why does it require so many games?

CallOn84 commented 5 months ago

It is not possible as a practical matter. Several hundred thousand games would be required for a net to begin to learn a particular style.

Why does it require so many games?

He might be referring to distil training, but I'm unsure. It would require a lot of games because it needs a lot of positions to learn and understand how to handle certain positions, and 1k games probably won't provide many positions for Leela to learn from.