Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags: [go, baduk, weiqi, katago, mcts, distillation]
|
| 4 |
+
---
|
| 5 |
+
# vibego-b6c96nbt-pat
|
| 6 |
+
|
| 7 |
+
561-MFLOP frontier point (0.9M params, ~8.2 CPU-ms) with the 3x3 dihedral pattern table (~+200 Elo at this size, ~0 FLOPs). Weights-only.
|
| 8 |
+
|
| 9 |
+
## Format / usage
|
| 10 |
+
|
| 11 |
+
PyTorch checkpoint: `{'model': state_dict, 'model_config': dict, 'spatial_subset': [...],
|
| 12 |
+
'global_subset': [...], 'step': int}` (full checkpoints also carry `optimizer`). Inputs are a
|
| 13 |
+
14-channel subset of KataGo v7 spatial features + 2 global features; the engine, training
|
| 14 |
+
pipeline, and evaluation harness will be released at https://github.com/sanderland/nanogo (the study writeup lives there
|
| 15 |
+
under `experiments/WRITEUP.md`). Strength numbers are judge scoreLead / win-rate Elo from
|
| 16 |
+
paired color-reversed-opening matches at 48 visits/move with a kata1-b18 judge.
|
| 17 |
+
|
| 18 |
+
Distilled from the public `kata1-b18c384nbt` net over katagoarchive.org positions — credit to
|
| 19 |
+
lightvector and the KataGo distributed-training contributors.
|