Evaluate with the NNUE network only

[fazega]

Is it possible to get the raw centipawn score from the pretained embedded neural network, for a given position?

[peregrineshahin]

Yes you can use the eval command to trace the NNUE evaluation of a position. but it does not tell too much compared to practice (combined with search) so probably a sample of x positions evaluated in such way can only tell you how root moves are sorted in the very first iteration without a TT.

See the example bellow where the net gives Final evaluation -0.02 (white side) [with scaled NNUE, ...] for the Scholar's mate !

Stockfish dev-20240121-a6fd17f2 by the Stockfish developers (see AUTHORS file)

position fen r1bqkbnr/ppp2ppp/2np4/4p3/2B1P3/5Q2/PPPP1PPP/RNB1K1NR w KQkq - 0 4
eval

info string NNUE evaluation using nn-baff1ede1f90.nnue
info string NNUE evaluation using nn-baff1edbea57.nnue


 NNUE derived piece values:
+-------+-------+-------+-------+-------+-------+-------+-------+
|   r   |       |   b   |   q   |   k   |   b   |   n   |   r   |
| -4.53 |       | -4.61 | -8.20 |       | -3.88 | -3.73 | -4.82 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   p   |   p   |   p   |       |       |   p   |   p   |   p   |
| -0.64 | -1.29 | -1.25 |       |       | -1.51 | -1.27 | -0.35 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |   n   |   p   |       |       |       |       |
|       |       | -4.20 | -1.07 |       |       |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |   p   |       |       |       |
|       |       |       |       | -1.46 |       |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |   B   |       |   P   |       |       |       |
|       |       | +4.93 |       | +1.77 |       |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |       |   Q   |       |       |
|       |       |       |       |       | +8.88 |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   P   |   P   |   P   |   P   |       |   P   |   P   |   P   |
| +0.54 | +0.89 | +1.25 | +0.76 |       | +0.79 | +1.40 | +0.60 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   R   |   N   |   B   |       |   K   |       |   N   |   R   |
| +3.62 | +3.06 | +3.40 |       |       |       | +3.03 | +3.89 |
+-------+-------+-------+-------+-------+-------+-------+-------+

 NNUE network contributions (White to move)
+------------+------------+------------+------------+
|   Bucket   |  Material  | Positional |   Total    |
|            |   (PSQT)   |  (Layers)  |            |
+------------+------------+------------+------------+
|  0         |  +  0.11   |  -  2.48   |  -  2.36   |
|  1         |  -  0.12   |  +  1.34   |  +  1.21   |
|  2         |  -  0.15   |  +  0.24   |  +  0.08   |
|  3         |  -  0.15   |  +  0.06   |  -  0.08   |
|  4         |  -  0.14   |  +  0.13   |  -  0.01   |
|  5         |  -  0.14   |  +  0.08   |  -  0.06   |
|  6         |  -  0.13   |  +  0.07   |  -  0.06   |
|  7         |  -  0.14   |  +  0.11   |  -  0.02   | <-- this bucket is used
+------------+------------+------------+------------+

NNUE evaluation        -0.02 (white side)
Final evaluation       -0.02 (white side) [with scaled NNUE, ...]