Rules engineering
Implement board state, legal actions, game end, and scoring, then define stable neural-network inputs and outputs. The approach is best suited to deterministic, two-player, perfect-information games and their variants.
Project / Board-game AI archive
KataGomo
A multi-game AI engineering project built from KataGo. It connects rule implementation, search, and self-play training, then publishes reproducible experiments as models, datasets, and release records.
Research scope
From a ruleset to a trainable, testable system
KataGomo studies the complete engineering path, not just the result of a single match. Built on KataGo, it extends AlphaZero-style methods to different boards, action spaces, and terminal conditions.
KataGomo 关注的是完整工程链路，而不只是某一次对局的结果。项目基于 KataGo，将 AlphaZero 方法扩展到不同棋盘、行动空间与终局条件。
Search and self-play
Reuse mature search and training infrastructure so a new ruleset can generate data through self-play. Each game still needs its exploration efficiency, sample distribution, and hyperparameters checked independently.
Models and datasets
Selected models, runnable releases, and datasets are public. Strength and win/loss statements on this site always belong to a specified ruleset, version, and evaluation setup; they are not unconditional claims across platforms or variants.
Published experiments
Games covered and the limits of the evidence
Public artifacts cover dozens of games and rule variants. Gomoku, Hex, Connect6, Ataxx, Animal Chess, and Chinese Checkers are among the areas with downloadable models or experiment records.
Game index
Open a game to find its models, version notes, and download links.
Interim findings
Some public models achieved strong playing results under the project's test settings. Experiments in Connect6, NoGo, Breakthrough, and other games also produced trends worth further verification.
For example, Connect6 self-play showed a first-player advantage, but whether that amounts to a forced win still depends on the exact rules, search budget, and stricter independent verification.
Cost and applicability
Small-network smoke tests can often reuse existing hardware, but full training has no universal price. Record the GPU model, GPU hours, self-play game count, and search budget, then estimate cost from the actual electricity or cloud-compute rate.
The project does not assume that implementing rules automatically produces a strong model; training stability and result quality still require case-by-case evaluation.
Rules experiments
Archive index
Code, models, data, and engineering notes
Choose the material by task. The download catalog centralizes published games instead of treating repository folders as documentation.
- Repository KataGomo main repository Source code, branches, and release records
- Catalog Games, models, and downloads Published artifacts for Gomoku, Hex, Connect6, and more
- Datasets ModelScope data profile Related public datasets
- Guide Training AI for a game you invented Branch selection, builds, checks, and training scale
- Reference C++ rules integration State, actions, terminal conditions, and network inputs
- Field note Engineering experience in board-game AI Search efficiency, model architecture, and training judgment
Release log
Recent updates
Recent public records as of July 13, 2026. KataGo network ratings change as game statistics accumulate; use the linked source for the current status.
| Date | Category | Record | Source |
|---|---|---|---|
| 3D connection games | Released the Lizzie3D interface and models for four three-dimensional connection rulesets. Ordinary 3D Gomoku was not included in this release. | 3D games release | |
| Go networks | The training site listed a b28c512nbt network as the latest upload; its strongest confidently rated network was a b40c768nbt version uploaded on June 6. | ||
| Ataxx | Released a new Transformer model. The release notes report a 200 Elo improvement over the preceding version. | Ataxx Release |