Olithink-uci interfaced with David Carteau's Cerebrum NNUE libraries for evaluation.
Olithink.nn trained on 30 million positions from over 3 million games from CCRL

New nets - stronger play.

Datasets created using a pgn file containing entire CCRL database of engine vs engine matches.
Matches containing duplicate moves were removed using pgn-extract. Pgn download > https://drive.proton.me/urls/VMFN932P4M#V0CNXOJKgQyD

Dataset/training was done on my 12 core Xeon computer in cpu mode - 24 hours to complete 11 epoch network files.

Would be a lot faster in gpu mode if I had a compatible nvidia gpu.

Last epoch, epoch 11 should be the best.

To created new olithink.nn file using the new network epochs: 

Delete existing olithink.nn and network.txt from olithink folder.

Rename epoch file you want to use to 'network.txt'

Olithink uses the smaller quantised files - the ones with 'q' in the name.

So for example rename 'epoch-11-q.txt' to 'network.txt'

Copy file to olithink folder and run olithink engine.

Olithink will detect missing 'olithink.nn' file and create a new one from 'network.txt'

If Olithink cannot find both 'olithink.nn' and 'network.txt' it will use traditional eval instead.

Uci hash option added

Added UCI configuratio option to address The NNUE "Halo Effect". The problem is Networks might look at a 
wild position and see a massive, non-material advantage (like a powerful attack) and completely ignore the fact that 
it has to give up a lot of pieces to get it. It believes its attack is so strong that the sacrifice is worth it, even 
if, deep down, it might not be, spotting subtle tactical and positional advantages, but can get a little too aggressive 
or "speculative." So I want to mix the material score with the NNUE score to create a good hybrid solution.

This modification implements a hybrid evaluation that mixes the NNUE score with the material score to mitigate the NNUE 
"Halo Effect". The material score is blended in at a ratio of 1/8, meaning the final score is (7/8 * Score_NNUE) + (1/8 * Score_Material)

Added analysis mode.

Fixed regression bugs.
Added nnue-hybrid ratios user-configurable in uci engine configuration screen