mirror of
https://github.com/official-stockfish/Stockfish.git
synced 2026-07-22 20:57:10 +00:00
This network is further trained on a new BT4 distillation stage, fine tuning on ~2 billion positions relabeled with the value head output of `BT4-tf13tune.pb.gz`. The dataset can be found at https://huggingface.co/datasets/xushawn/test80-bt4-relabel. A modified branch of lc0 was used to derive this data: https://github.com/xu-shawn/lc0/tree/relabel_dual_stream_test 2 billion positions represent a tiny subset of the total training data, and BT4 relabeling is inherently computationally expensive. I expect a lot more gains as more data are relabeled, but it will likely require coordinated community effort. Everyone is welcome to contribute, and yl25946 has made a spreadsheet to track progress: https://docs.google.com/spreadsheets/d/1yanofhusEzDg8ZnurAw799ikoTY6GcqsNMYfpswOIbw/edit. Special thanks to Viren6, who performed policy/value distillation experiments on Monty, and created the lc0 distillation fork that the current relabeler is based on; yl25946 for proposing the idea of large network distillations back in February 2025, running distillation experiments on the HL4096 network, and working on fine tuning attempts; vondele for nettest and suggesting the fine-tuning approach; and many others on the knowledge distillation thread in the SF Discord #ideas channel. nettest PR: https://github.com/vondele/nettest/pull/369 Ongoing STC: LLR: -0.01 (-2.94,2.94) <0.00,2.00> Total: 72224 W: 18891 L: 18784 D: 34549 Ptnml(0-2): 336, 8437, 18332, 8798, 209 https://tests.stockfishchess.org/tests/view/6a3ae7913036e45021aeb4a0 Passed LTC: LLR: 2.94 (-2.94,2.94) <0.00,2.00> Total: 25110 W: 6566 L: 6288 D: 12256 Ptnml(0-2): 27, 2625, 6957, 2935, 11 https://tests.stockfishchess.org/tests/view/6a3b73513036e45021aeb51e Passed VLTC: LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 18544 W: 4924 L: 4658 D: 8962 Ptnml(0-2): 5, 1730, 5533, 2002, 2 https://tests.stockfishchess.org/tests/view/6a3bbe233036e45021aeb56e closes https://github.com/official-stockfish/Stockfish/pull/6924 Bench: 2710209 Co-authored-by: Li Ying <121075683+yl25946@users.noreply.github.com> Co-authored-by: Viren6 <94880762+Viren6@users.noreply.github.com>
56 lines
1.7 KiB
C++
56 lines
1.7 KiB
C++
/*
|
|
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
|
|
Copyright (C) 2004-2026 The Stockfish developers (see AUTHORS file)
|
|
|
|
Stockfish is free software: you can redistribute it and/or modify
|
|
it under the terms of the GNU General Public License as published by
|
|
the Free Software Foundation, either version 3 of the License, or
|
|
(at your option) any later version.
|
|
|
|
Stockfish is distributed in the hope that it will be useful,
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
GNU General Public License for more details.
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
along with this program. If not, see <http://www.gnu.org/licenses/>.
|
|
*/
|
|
|
|
#ifndef EVALUATE_H_INCLUDED
|
|
#define EVALUATE_H_INCLUDED
|
|
|
|
#include <string>
|
|
|
|
#include "types.h"
|
|
|
|
namespace Stockfish {
|
|
|
|
class Position;
|
|
|
|
namespace Eval {
|
|
|
|
// The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
|
|
// for the build process (profile-build and fishtest) to work. Do not change the
|
|
// name of the macro or the location where this macro is defined, as it is used
|
|
// in the Makefile/Fishtest.
|
|
#define EvalFileDefaultName "nn-f8a759c05f9f.nnue"
|
|
|
|
namespace NNUE {
|
|
class Network;
|
|
struct AccumulatorCaches;
|
|
class AccumulatorStack;
|
|
}
|
|
|
|
std::string trace(Position& pos, const Eval::NNUE::Network& network);
|
|
|
|
Value evaluate(const NNUE::Network& network,
|
|
const Position& pos,
|
|
Eval::NNUE::AccumulatorStack& accumulators,
|
|
Eval::NNUE::AccumulatorCaches& caches,
|
|
int optimism);
|
|
} // namespace Eval
|
|
|
|
} // namespace Stockfish
|
|
|
|
#endif // #ifndef EVALUATE_H_INCLUDED
|