Data provenance, training pipeline and results for the NNUE networks of Luna Chess Engine — not to be confused with Thomas Mergener's Luna, a different engine.
The network in Luna CE
v4.0.0
(resources/net.bin, sha256
ecd8a917fd8d6f1e900f64ec1721494178cd610ab5f02b4c91c0ca1950446940): trained
by the author, with his own training pipeline, on public Leela Chess Zero
data (phase-1 line, not the gen1/gen2/gen3 line below), on 1 billion
positions (about 914 million unique by exact board and side-to-move match,
91.4%; 910.5 million if mirror images are merged). Measured in the engine's
own search against akimbo's network: statistically indistinguishable,
+0.3 ± 10.7 Elo (95% CI, 2,000 games) — parity, not a win. Full
provenance: PROVENANCE.md. Data source and license: DATA.md.
Result detail: results/RESULTS_v4.0.0.md. Traps hit along the way:
LESSONS.md.
The network shipped in v4.0.0 comes from a separate line of work (the
phase-1 pipeline, see results/) from the fully self-play
gen1/gen2/gen3 line described below, which is a different, ongoing
effort — nothing from that line has been shipped yet.
Every label used to train these networks was produced by Luna's own search,
and every position comes from Luna's own self-play. The one external
influence anywhere in the chain — Stockfish filtering gen1's opening pool, a
selection of starting positions and not a label — is documented, not hidden.
LINEAGE.md traces each network back to its ancestors, COMPLIANCE.md
states how each stage of the pipeline meets that claim, and RESULTS.md holds
the measurements — with the raw per-position data in results/, so every
number here can be recomputed.
This repository holds the history of how each network was made and measured. The engine's source code lives in its own repository.
| File | Covers |
|---|---|
PROVENANCE.md |
v4.0.0 network: trainer, architecture, config, machine, quantisation gates |
DATA.md |
v4.0.0 training data: source and license, two levels (Lc0 games / Linrock's conversion) |
results/RESULTS_v4.0.0.md |
v4.0.0 network: match results table |
LESSONS.md |
Traps found during the phase-1 line, with evidence |
LINEAGE.md |
gen1/gen2/gen3 line: ancestry of each network |
COMPLIANCE.md |
gen1/gen2/gen3 line: how each stage meets the self-play-only claim |
RESULTS.md |
gen1/gen2/gen3 line: measurements |
The numbers are not always flattering: the compliant network currently ranks
below gen0 (trained on externally produced labels, kept only as a reference
point) and below the third-party akimbo network (used as a yardstick, never
to generate data). They are published anyway. A record that shows the
inconvenient numbers is worth more than a provenance claim nobody can check.