Go player and era models
16 models: 13 unique players and 3 eras, all in this repository.
One selected released model per player or era. Updated 18 November 2026 to
include the five players previously available only in go_imitation_weights.
This repository is public. The game-record dataset remains private.
Models with standardized development evaluations
These eleven models retain their previously released weights and evaluation records. Their GTP examples now explicitly specify the logging settings required by KataGo.
| 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 five main player datasets use at least 500 training games. Ke Jie, Ma Xiaochun and Ran are smaller pilot datasets. For these November development models, selection used the lowest recorded full-validation NLL among comparable candidates; test scores did not determine selection. Go Seigen uses direct adaptation, and Ke Jie uses final-block adaptation.
Consolidated legacy player models
| Player | Original released directory |
|---|---|
| Cho Chikun | cho-chikun_export |
| Kitani Minoru | kitani_export |
| Nie Weiping | nieweiping_export |
| Honinbo Shusaku | shusaku_export |
| Takemiya Masaki | takemiya_export |
These five checkpoints and native exports were copied byte-for-byte from
go_imitation_weights.
Their original training statistics and source fingerprints are included.
Standardized held-out performance and game-level split counts for these exact
legacy exports have not been established here, so no comparable test scores
are claimed. Per-model cards distinguish them from the newer evaluations.
The 13 players are Go Seigen, Sakata Eio, Kobayashi Koichi, Cho Hunhyun, Kato Masao, Ke Jie, Ma Xiaochun, Ran, Cho Chikun, Kitani Minoru, Nie Weiping, Honinbo Shusaku and Takemiya Masaki. Cho Chikun and Cho Hunhyun are distinct players. The three era models use 1950, 1970 and 1990 profiles.
Each model has model.ckpt and model.bin.gz: 32 weight files representing
16 models, not 32 models. No intermediate checkpoints, alternative training
recipes, style adapters, or multi-move/RL experimental variants were added.
The recent controlled study uses different data splits and does not establish
that its checkpoints should replace these releases.
Download and run
from huggingface_hub import snapshot_download
folder = snapshot_download(
repo_id="fastwalker1118/go-player-imitation-checkpoints",
allow_patterns=["models/cho-chikun/*", "model_catalog.json"],
token=False,
)
From a downloaded model directory:
katago gtp -model model.bin.gz -config gtp_example.cfg
Use a compatible KataGo build. Native checks use KataGo 1.16.3.
Preserve the supplied metadata profile and full history; match rules and komi
to the game. Legacy per-model cards identify illustrative profiles. Nie Weiping's
legacy network has no metadata encoder. GTP search behavior is not identical to
raw-policy evaluation. Python checkpoints use KataGo's loader, not Transformers.
November development models used KataGo Python revision
a1d832c693dbed9026bf56552645f824f1c6ece5; this is not asserted as the original
legacy training revision. Original checkpoints can contain training state and
legacy local-path metadata, which are not needed for native inference.
Interpretation and data
Top-1 is recorded-move agreement, not win rate or Elo. NLL is lower-is-better. Different players and experiment generations use different test sets; their numbers are not a controlled cross-model ranking. Local fine-tuning holdouts do not rule out upstream pretraining overlap. Era adaptation did not improve test NLL in the November development runs. The Ma Xiaochun pilot contains only games against Nie Weiping. No tournament-strength result is claimed.
The November development models start from KataGo's Human SL network. The older Nie Weiping model comes from the legacy all-human network without a metadata encoder; source details for legacy releases are preserved without inferring missing training settings. See per-model cards and metadata.
The private dataset remains private. No SGF records, feature datasets or private dataset files are included here. Model catalog, data catalog, and checksums document the collection.
Collection history
The main collection previously contained 8 player models plus 3 era models. Five existing player releases were consolidated here on 18 November 2026, bringing this repository to 13 players plus 3 eras. The old repository remains available as an archive; its five models duplicate those now included here and must not be counted again as five additional distinct models.
Attribution
Derived from KataGo and its published
models, including the Human SL network.
Upstream code attribution is retained in KATAGO_LICENSE.txt. No new public
redistribution license for datasets or derived weights is asserted. Public model
access does not publish or grant access to the underlying private game records.