SPCC: Testrun of Fire 8.N Raptor finished

[kranium]

The "Raptor" nn.bin network is not original either. It is just one of SV's earlier Stockfish networks: https://tests.stockfishchess.org/api/nn ... fd0df.nnue

These networks are not 'Stockfish' networks...but I do understand your need to categorize them as such, making it seem as if anyone using them was taking from SF.

Some or many of Sergio's networks were certainly tested in Fishtest, but the vast majority (or all as far as I know including the one that I call Raptor) did not succeed, and did not become a default network.

To anyone interested, or to see if any particular SV network was tested by SF, or used elsewhere...all his networks are available here:
https://www.comp.nus.edu.sg/~sergio-v/nnue/

More than fair to call a network trained with Stockfish evals, from Stockfish games, in a trainer wrapped around Stockfish, a Stockfish Network

Apparently you're much more familiar with his methods than I am..
but if someone is trying to create the strongest network possible, NOT using SF games is certainly a mistake IMO

[connor_mcmonigle]

...
Apparently you're much more familiar with his methods than I am..
but if someone is trying to create the strongest network possible, NOT using SF games is certainly a mistake IMO

I can second what Andrew is saying and if you're still inclined to doubt I can point you to SV's explanation of his methodology.

Sure, if you're only interested in chasing Elo, then best would be to train on SF games, but then you'd also be best off not writing your own engine and just forking Stockfish to begin with in my opinion... Most testers and tournaments have stated that they are largely uninterested in testing engines with networks trained on Stockfish generated data, using training code written by Stockfish developer and using inference code derived from Stockfish/CFish, viewing such engines as fairly derivative.