vibego-s11-b12c152nbt-pat

OVERALL STUDY CHAMPION. b12c152nbt-pat (4.21M params, 2686 MFLOP/eval, ~30.8 single-thread CPU-ms). 260k steps on the full pool (44M kata1 positions + weak-era x3, ~20% mix). Strength: BEATS g170e-b10c128 by +7.7 ± 2.4 judge scoreLead [+3.0, +12.5] over 192 paired games at 48 visits (b18 judge, 256v); h2h vs the 2.6M s10 +21.0 ± 7.8. The b10 tier was cleared by a capacity step — same data as s10 (which measured −9.5 sL), +17 sL from width alone. Full checkpoint (optimizer included, resume-capable).

Format / usage

PyTorch checkpoint: {'model': state_dict, 'model_config': dict, 'spatial_subset': [...], 'global_subset': [...], 'step': int} (full checkpoints also carry optimizer). Inputs are a 14-channel subset of KataGo v7 spatial features + 2 global features; the engine, training pipeline, and evaluation harness will be released at https://github.com/sanderland/vibego (the study writeup lives there under experiments/WRITEUP.md). Strength numbers are judge scoreLead / win-rate Elo from paired color-reversed-opening matches at 48 visits/move with a kata1-b18 judge.

Distilled from the public kata1-b18c384nbt net over katagoarchive.org positions — credit to lightvector and the KataGo distributed-training contributors.