25x25, KataGo (B) - Sayuri
KataGo b28-s728 playouts 5
Sayuri b20 2385k playouts 500

25x25, KataGo (B) - Sayuri
KataGo b28-s716 playouts 5
Sayuri b20 uec16 playouts 500

Hi

You'll find here a 50x50 selfplay game by KataGo: enjoy!

KataGo now plays well on this board size. While there are still occasional scoring errors (particularly for boards larger than 40x40), overall performance is solid

I’ll share more games in the coming weeks and plan to release the model (.npz file) in the coming months.

A huge thanks to lightvector for KataGo, and for his guidance in setting up the selfplay loop for large boards!

What a stunning surprise!!! which version of Katago are you using? Can you upload it?

Important precision: it's a specific 28b KataGo model I trained for more than one year on large boards, up to 50x50, starting from the latest available kata1 (I began with the latest 18b end of 2023, then switched to 28b in 2024).
More precisions during the weekend : -)

Some more info: this "large board" KataGo network has been trained in selfplay since late 2023. Initial training used an 18b network (based on the latest kata1 run available 18b at that time).
In September 2024, I switched to 28b and immediately noticed a significant strength boost on all board sizes, including large boards —prompting a full transition to 28b for both selfplay and training.
Again, a huge thanks to lightvector for KataGo, and for his guidance in setting up the selfplay loop for large boards!

Training Process / selfplay mix:
- Around 85% of the data from upscaled kata1 selfplay (adjusted to 50x50 tensor sizes)
- Around 15% of the data from native large-board selfplay (generated directly on 50x50 tensors)

Hardware: 3 GPUs (RTX 4090, 4080, and 3090) running continuously for over a year.

Selfplay Evolution:
- Games generated up to 29x29 in 2023.
- From early 2024: expanded to 50x50, after resolving some stability issues. Proportion of large board games was progressively increased to 32x32 → ~40x40 → 50x50, also including rectangular boards (e.g., 7x50) like kata1.

Performance Insights:
Kata1 (« the main distributed run ») Baseline:
- ≤25x25: Strong native play.
- ≤30x30: Occasional errors (e.g., premature passes/first-line moves).
- 30x30 to 40x40: Frequent errors; above 40x40 requires quasi-systemic manual supervision to avoid frequent major mistakes or refusal to play (pass).
- Note : surprisingly, kata1 natively plays quite well on “very rectangular” boards such as 7x50. So the LB-trained network is better, but not insanely better.

LB18b (Large Board 18b, from 2023 to August 2024):
- Rapid (but limited) improvement up to 29x29, slower (but much bigger) progress beyond 35x35.

LB28b (Large Board 28b):
- Trained initially using LB18b selfplay data: it outperformed LB18b on all board sizes
- Then trained natively from LB28b selfplay data.

Quantitative Results (LB28b vs k1-s7503, 400 games, 1000 visits):
- 19x19: ~25 Elo gain over k1-s7503 (boost probably due to a lower learning rate).
- 25x25: around +200 Elo.
- 29x29: around +600 Elo (395 wins, 1 draw, only 9 losses out of 405 games).
- ≥30x30: LB28b largely dominates k1-s7503.

Progress Metrics on large boards (vs earlier LB-trained networks):
LB18b (2024):
- 37x37: +400 Elo (25 visits, April 2024 → Sept 2024).
- 50x50: +700 Elo (25 visits, June 2024 → Sept 2024).
LB28b (Sept. 2024 - March 2025):
- 37x37: +150 Elo (1K visits, Sept 2024→March 2025).
- 50x50: +350 Elo with 10 visits and +900 Elo with 100 visits (398 wins, 2 losses !).

Current Limitations (LB28b):
- Scoring errors: some minor discrepancies (≤few points) on boards >37x37, some rare strong discrepancies
- Dame issues: Residual neutral points on boards mainly above 40x40.

Next Steps:
- Uploading LB28b vs kata1 29x29 games, and some 50x50 games in the coming days!
- More large-board games to follow, with a few high visits large board games.
- Releasing the network after (probably) a last training by mid-April

And attached, you'll find 8 games from the 29x29 match against kata1 at 1000 visits (with about +600 Elo, cf quantitative results above). It's a txt file that you can split in 8 different sgf files if needed.

is the Katago version available for download now for boards larger than 29x29 and 28b nets?

I'm using the TRT version, which works. I did not test the other versions recently, not sure anymore.
They should be "clean", but please check again before using them.
And IIRC, I compiled them for 50x50 or 51x51 board sizes.

katagoOCL:

My bad, the OpenCL version was not the good one, I modified the 7zip file (and the OCL version is now called OCL50): it should work hopefully!

thank you very much! It's working! b28-s832 on board 51 of course "at a loss"

37x37 Katago - Katago, b28-s844, 200 playouts

Here is the first game I generated on a 50x50 games played with 10K visits!
It is between the two most recent networks trained on large boards (which are quite comparable in strength).

On 33x33 board, some games of KataGo's network trained on large boards vs latest standard kata1 network (kata1-s8446), with H5 and 10 times less visits

an error with the placement of handicap stones

It's not an error, it's a choice made by lightvector to generate more diversity in handicap games: the handicap stones are placed "randomly".

it would be interesting to play with this network. Can you upload it?

See above my post and the "next steps" : I'm releasing it in the coming weeks, I'm currently generating some more selfplay and will do some more training tests before releasing the best network ; -).

27x27 Katago - Katago, b28-fd3 - b28-s844
maxPlayouts = 50