Details about the training data of BT3/BT4? #1981
Unanswered
jiacheng-ye
asked this question in
Q&A
Replies: 1 comment
-
|
BT3 and BT4 were trained using data generated by our training pipeline for T80. I'm not sure of the exact game count, it was using what was created over about 5 months. Our latest training data are in https://storage.lczero.org/files/training_data/test80/ and their format is described in https://github.com/LeelaChessZero/lc0/blob/master/src/trainingdata/trainingdata.h. For mode details join our chat: https://discord.gg/pKujYxD |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
hey any planning into beating stockfish anytime soon ? |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi there, thanks for the great work! Based on the recent blog, I'm curious about the data used when training BT3 and BT4. Are they the same as in the DeepMind paper (i.e., 10M games reannotated by stockfish 16), and whether they are available?
Beta Was this translation helpful? Give feedback.
All reactions