How to Contribute to KataGo Distributed Training

To contribute, download a supported KataGo client (v1.18.0 or later is recommended), run it with your own configuration, and let it generate self-play games that are uploaded to the distributed training run. The site's "How to Contribute" section is the authoritative starting point, and the "Downloading KataGo" section provides the client. Contributions are what produce the training data used to improve the neural network.

What the project is

KataGo is an open-source, self-play-trained Go engine. According to the site, it can predict score and territory, play handicap games reasonably, and handle many board sizes and rules with the same neural net. This site hosts KataGo's first public-distributed training run, attempting to resume training from the end of the previous official run ("g170") that ended in June 2020.

Before you start

  • Client version: Support for versions prior to v1.18.0 has been retired (update posted 2026-09-15). Use v1.18.0 or later.
  • Why version matters: v1.18.x added performance optimizations for transformer models and rules fixes for data generation. The run has switched to transformer models for self-play, so older clients are no longer accepted.
  • Hardware: The site does not state minimum hardware requirements in the material available here. Expect to need a machine capable of running KataGo's neural net inference; check the client's own documentation for specifics.
  • Account: The site shows "Log In" and a registration link, so contributing likely involves an account. The exact registration and login flow is not described in the available page evidence — follow the on-site instructions.

Steps to contribute

  1. Read the "How to Contribute" section on katagotraining.org. This is where the project's own instructions live, and they take precedence over any general advice.
  2. Download the client from the "Downloading KataGo" section, or from the linked KataGo GitHub release page.
  3. Install a supported version — v1.18.0 or later. If you already run an older version, upgrade before contributing.
  4. Configure and run the client so it performs self-play and submits games. The site's contribution instructions cover the configuration; the client generates the games and uploads them.
  5. Verify your contributions are counted. The site publishes stats (for example, "Stats for kata1") covering networks, games, and contributions, so you can confirm your client is submitting successfully.
  6. Keep the client updated. The project has repeatedly retired older versions once enough contributors migrated, so staying current avoids sudden rejection of your submissions.

What your contribution produces

Your client's self-play games become training data. The project's updates describe this pipeline directly: new models are trained and then used for self-play and rating games. For example, the 2026-04-08 update notes that a large model, "kata1-zhizi-b40c768nbt-fdx6c," was made available on the networks page and started being used for self-play, with its training data to be mirrored on the KataGo data archive site. The 2026-08-25 update notes that the first transformers were uploaded to the run and self-play was switching to them.

Common pitfalls

  • Running an outdated client. Versions before v1.18.0 are no longer supported, so games from them will not be accepted.
  • Upgrading late. The project disables support for older versions once enough contributors have migrated. Upgrading promptly keeps you contributing without interruption.
  • Assuming contributions are only about raw compute. The updates show the project also values testing and optimization help from community members, alongside donated self-play.

Where to check status

  • Updates on the site for version requirements and run changes.
  • Stats for kata1 for networks, games, and contributions.
  • Networks page for current models, and the linked data archive for training data mirrors.
katagotraining.org