Go player and era models
11 models: 8 players and 3 eras. One final selected model per player or era, with no alternative recipes, interpolations, intermediate checkpoints, or style adapters in the current collection. This model repository is public. The game-record dataset remains private and is not included in this repository. Collection simplified on 17 November 2026; experiments were completed 12 November.
Models
| Model | Type | Training games | Test games | Exact move match | Move NLL |
|---|---|---|---|---|---|
| Go Seigen | player | 500 | 105 | 58.877% | 1.318694 |
| Sakata Eio | player | 1051 | 213 | 59.355% | 1.301686 |
| Kobayashi Koichi | player | 1345 | 314 | 61.697% | 1.199112 |
| Cho Hunhyun | player | 855 | 176 | 62.499% | 1.193499 |
| Kato Masao | player | 1151 | 250 | 62.369% | 1.178779 |
| Ke Jie | player | 47 | 10 | 59.118% | 1.253225 |
| Ma Xiaochun | player | 59 | 12 | 57.536% | 1.332147 |
| Ran | player | 174 | 37 | 60.183% | 1.361987 |
| 1950 era | era | 600 | 100 | 59.239% | 1.321798 |
| 1970 era | era | 599 | 100 | 59.410% | 1.306577 |
| 1990 era | era | 598 | 100 | 59.322% | 1.317459 |
The eight players are Go Seigen, Sakata Eio, Kobayashi Koichi, Cho Hunhyun, Kato Masao, Ke Jie, Ma Xiaochun, and Ran. Era models use 1950, 1970, and 1990 metadata profiles. The five main player datasets have at least 500 training games; Ke Jie, Ma Xiaochun, and Ran are smaller pilot datasets.
For players with multiple trained versions, the retained model has the lowest recorded full-validation NLL. Test scores did not determine this choice. Go Seigen uses the direct-adaptation checkpoint; Ke Jie uses the final-block-adaptation checkpoint. Their source run names remain in metadata for reproducibility.
Each directory contains model.ckpt and its matching model.bin.gz native
KataGo export: two formats of one model, giving 22 weight files for 11 models.
Model catalog and checksums record
the selected models and their provenance.
Download and run
Download without authentication:
from huggingface_hub import snapshot_download
folder = snapshot_download(
repo_id="fastwalker1118/go-player-imitation-checkpoints",
allow_patterns=["models/cho-hunhyun/*", "model_catalog.json"],
token=False,
)
From the downloaded models/cho-hunhyun directory:
katago gtp -model model.bin.gz -config gtp_example.cfg
Use a compatible KataGo build. Native validation used version 1.16.3. Preserve
the provided profile and full history; match rules and komi to the actual game.
PyTorch checkpoints use KataGo's loader, not Transformers. The corresponding
KataGo Python source revision is a1d832c693dbed9026bf56552645f824f1c6ece5.
Interpretation and data
Top-1 is recorded-move agreement, not win rate or Elo. Move NLL is lower-is-better. No tournament-strength result is claimed. Holdouts are disjoint from local fine-tuning; overlap with upstream pretraining remains unknown. Players have different test sets, so cross-player accuracy is not a controlled model ranking. Era adaptation did not improve test NLL in these runs. The Ma Xiaochun pilot has only games against Nie Weiping. The value head was not directly trained, although trunk updates can alter value predictions.
The private dataset contains the exact
curated game records and NPZ features. The upstream initialization was the
published b18c384nbt-humanv0.ckpt Human SL model. It is not duplicated here.
Each model directory includes its training configuration and evaluation summary.
Original checkpoint bytes retain legacy local-path strings in metadata; those
paths are not needed to load the weights.
Account model count
This collection adds 11 models to the 5 older player models in
go_imitation_weights:
16 Go models total β 13 players and 3 eras. Robotics checkpoints and empty
repositories are not included in that Go-model count.
Attribution
Derived from KataGo and its published
Human SL network.
Upstream code attribution is preserved in KATAGO_LICENSE.txt. No new public
redistribution license for datasets or derived weights is asserted. The model
repository was made public at the owner's request on 17 November 2026; this does
not publish or grant access to the underlying private game-record dataset.