vibego
Collection
Tiny distilled KataGo-style Go nets (one GPU, public data) + the distillation datasets. Writeup in the repo. • 14 items • Updated
vibego-s9-b10c128nbt-pat
b10c128nbt-pat (2.57M params, 1567 MFLOP/eval, ~17.8 single-thread CPU-ms). 240k steps on 23.5M positions (kata1-2400sh + full weak-era pool). −8.6 ± 2.4 judge scoreLead vs g170e-b10c128 over 192 paired games (48 visits, b18 judge 256v) at 0.7x the anchor's FLOPs; decisively above g170-b6c96. Full checkpoint (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.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support