Hugging Face's logo

Datasets:
sanderland
/
vibego-replay-relabel-ab

Tags:
go
baduk
weiqi
katago
distillation
License:
Dataset card Files
xet
 Community
Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

nanogo-replay-relabel-ab

Paired label-quality A/B sets: the SAME 450k replayed g170 positions (full move-history context, real history input planes) labeled two ways — v1/ b18 raw prior at 1 visit, v32p/ b18 32-visit searched policy (visit-count distribution) + searched value. Study result: searched labels only reach parity with the raw prior at 32x label cost (and are decisively harmful if relabeling lacks history context).

Format

Each .npz holds training rows in KataGo-v7-compatible arrays:

  • binaryInputNCHWPacked (N, 22, 46) uint8 — packed 19x19 binary input planes (v7 channel order)
  • globalInputNC (N, 19) float32 — global features (komi at index 5 = selfKomi/20, side-to-move relative)
  • policyTargetsNCMove (N, 1, 362) float32 — teacher policy target (361 moves + pass), to-move
  • globalTargetsNC (N, 64) float32 — [0:3] win/loss/nr, [3] scoreLead (to-move), [27] ownership weight
  • valueTargetsNCHW (N, 1, 19, 19) — ownership target, to-move-positive

Teacher: public kata1-b18c384nbt queried at the stated visits. Upstream positions from katagoarchive.org (g170 self-play / kata1 daily training data) — credit to lightvector and the KataGo distributed-training contributors. License: CC-BY-4.0.
Downloads last month
10
Total file size:
1.51 GB

Collection including sanderland/vibego-replay-relabel-ab