19 Oct 21:38

0c509ae

Latest

🏆 KW29 Custom Fine‑Tuned Model Release: kw29‑b18c384nbt‑1018

Release Date: 2025‑10‑19
Author: changcheng967
With support from: Doulet Media
Powered by: KataGo — full credit to the KataGo authors, license, and required attributions.

📦 Overview

This release introduces KW29 (b18c384nbt‑1018), a custom fine‑tuned KataGo model.
It has been trained on ~2 million samples, reaching a scale of ~10 billion parameters.
The evaluation benchmark compares KW29 against the baseline FMSWA7 (b28c512) under equal‑time conditions.

🧪 Experiment Setup

Board size: 19x19
Baseline: FMSWA7 (b28c512)
KW29: b18c384nbt‑1018
Visits: FMSWA7 = 10, KW29 = 40 (equal‑time fairness)
Games played: 100
Resignation: enabled (threshold −0.95, 6 turns)
Backend: TensorRT on RTX 5080 Laptop GPU

📊 Results Summary

Match Log Metrics

Model	Avg Move Time (s)	Total Moves (sample)	NN Rows
FMSWA7	0.0754	8068	387,146
KW29 b18c384nbt	0.0621	8087	633,603

➡️ KW29 is ~21% faster per move than FMSWA7 under equal‑time settings.

Elo & Win Rates

Model	Win% vs Opponent	Elo (± error)
FMSWA7	70.5%	+48.41 ± 25.16
KW29 b18c384nbt	29.5%	−48.41 ± 25.16

➡️ FMSWA7 remains stronger overall, but KW29 shows improved throughput and efficiency.

Derived Metrics

Model	Relative Speed	Elo / NN Row	Efficiency (Elo/sec)	Verdict
FMSWA7	1.00	+1.25e‑4	+641	Baseline
KW29 b18c384nbt	1.21	−7.6e‑5	−780	Weaker but more efficient

📈 Key Insights

Training scale: KW29 is trained with ~2M samples, totaling ~10B parameters.
Efficiency: KW29 processes significantly more NN rows, achieving higher throughput.
Strength: FMSWA7 still leads in Elo, but KW29 narrows the gap compared to earlier KW versions.
Fairness: Visit scaling (10 vs 40) produced nearly identical time per move, ensuring equal‑time fairness.

✅ Next Steps

Continue training KW series (KW30+) to close the Elo gap.
Track Elo progression across versions with the KW Report dashboard.
Explore tuning resign thresholds and temperature for further balance.

🙌 Credits

Created by changcheng967 with help from Doulet Media
Built on the open‑source KataGo engine

Full Changelog: KW28-b18c384nbt-1017...KW29-b18c384nbt-1018

Assets 3

18 Oct 15:36

changcheng967

KW28-b18c384nbt-1017

20da6ca

KW28-b18c384nbt-1017

Full Changelog: KW20-b28c512nbt-0930...KW28-b18c384nbt-1017

Assets 3

17 Oct 18:39

changcheng967

KW27-b18c384nbt-1017

c227348

KW27-b18c384nbt-1017

Full Changelog: KW19-b28c512nbt-0930.bin...KW27-b18c384nbt-1017

Assets 3

17 Oct 18:34

changcheng967

KW20-b28c512nbt-0930

c227348

KW20-b28c512nbt-0930

Full Changelog: KW19-b28c512nbt-0930.bin...KW20-b28c512nbt-0930

Assets 3

16 Oct 00:03

changcheng967

KW19-b28c512nbt-0930.bin

40eeaa9

KW19-b28c512nbt-0930.bin

What's Changed

Revert changes and update training branch by @changcheng967 in #23
Update results.xlsx with new data by @changcheng967 in #27
Update training branch by @changcheng967 in #29
Update Training branch by @changcheng967 in #32
Training by @changcheng967 in #33

Full Changelog: kw17-b28c512nbt-0928...KW19-b28c512nbt-0930.bin

Contributors

changcheng967

Assets 3

30 Sep 01:36

changcheng967

kw18-b28c512nbt-0929

436efd5

kw18-b28c512nbt-0929

🏆 kata_web Custom Fine-Tuned Model Release: kw18-b28c512nbt-0929

🔗 Trained using the KataGo open-source framework — full credit to KataGo authors, license, and required attributions.

📌 Overview

kw18 is a fine-tuned 19x19 Go model based on the KW17 checkpoint (kw17-b28c512nbt-0928).
It was trained for 1 epoch on 200,000 samples, totaling 3,050,944 training steps. KW18 continues the KW-series with the strongest Elo so far, achieving a 55% win rate and a new peak in Calculated Elo.

🧠 Model Information

Attribute	Value
Model Name	kw18-b28c512nbt-0929.bin
Configuration	b28c512nbt (28 blocks, 512 channels)
Board Size	19x19
File Size	~330 MB
Base Model	kw17-b28c512nbt-0928
Training Steps	3,050,944
Training Data	200k rows × 1 epoch of shuffled self-play
Training Time	~1.5 hours (A100-class GPU, approximate)
Framework	KataGo v1.17.0+ (PyTorch export)

📊 Performance Metrics

Model	Games	Wins	Loses	Win rate (%)	Avg Time (s)	Total Playouts	Baseline Elo	Elo Diff	Calculated Elo	Rank	Weighted Elo	Samples	Elo Realistic	Efficiency
KW18-b28c512nbt-0929	20	11	9	55	389.530	217,964	14085	+35.22	14120.22	1	14068.31479	3,050,944	4663	559.5563885

⚡ KW18 surpasses KW17 with a 55% win rate, higher Calculated Elo (14120), and best-in-series Weighted Elo. This marks a new performance peak for the KW line.

🚀 Usage Instructions

1. Download the model

wget https://github.com/changcheng967/kata_web/releases/download/v1.0/kw18-b28c512nbt-0929.bin
mv kw18-b28c512nbt-0929.bin ~/KataGo/models/

2. Download checkpoint (optional)

wget https://github.com/changcheng967/kata_web/releases/download/v1.0/model.ckpt
mv model.ckpt ~/KataGo/models/

3. Run with GTP

./cpp/main gtp \
  -model ~/KataGo/models/kw18-b28c512nbt-0929.bin \
  -config cpp/configs/gtp.cfg

Example session:

boardsize 19
clear_board
genmove B

4. GUI Integration

Sabaki → add engine: ./cpp/main, args: gtp -model models/kw18-b28c512nbt-0929.bin
Lizzie → configure with same GTP arguments
KaTrain → add as custom GTP engine

📦 Files Included

File	Purpose
kw18-b28c512nbt-0929.bin	Final exported model
model.ckpt	Checkpoint file for resuming training
results.xlsx	Full benchmark results (KW10 → KW18)

Full Changelog: kw16-b28c512nbt-0928...kw18-b28c512nbt-0929

Assets 5

30 Sep 01:10

changcheng967

kw17-b28c512nbt-0928

436efd5

kw17-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw17-b28c512nbt-0928

🔗 Trained using the KataGo open-source framework — full credit to KataGo authors, license, and required attributions.

📌 Overview

kw17 is a fine-tuned 19x19 Go model based on the KW16 checkpoint (kw16-b28c512nbt-0928).
It was trained for 1 epoch on 200,000 samples, totaling 2,840,000 training steps. Compared to KW16, it stabilizes performance and achieves a balanced 50% win rate, with stronger Elo and improved consistency.

🧠 Model Information

Attribute	Value
Model Name	kw17-b28c512nbt-0928.bin.gz
Configuration	b28c512nbt (28 blocks, 512 channels)
Board Size	19x19
File Size	~330 MB
Base Model	kw16-b28c512nbt-0928
Training Steps	2,840,000
Training Data	200k rows × 1 epoch of shuffled self-play
Training Time	~1.4 hours (A100-class GPU, approximate)
Framework	KataGo v1.17.0+ (PyTorch export)

📊 Performance Metrics

Model	Games	Wins	Loses	Win rate (%)	Avg Time (s)	Total Playouts	Baseline Elo	Elo Diff	Calculated Elo	Rank	Weighted Elo	Samples	Elo Realistic	Efficiency
Baseline (FMGo7)		0	14085	2	11268		14085						4660
KW14-b28c512nbt-0927	20	7	13	35	361.964	202,978	14085	-120.41	13964.59	7	13919.19557	1,240,000	4647	560.7684742
KW15-b28c512nbt-0928	20	9	11	45	333.465	186,586	14085	-38.38	14046.62	5	13994.8897	1,940,000	4656	559.5369829
KW16-b28c512nbt-0928	20	7	13	35	364.026	207,507	14085	-155.63	13929.37	8	13929.37	2,640,000	4643	570.0334591
KW17-b28c512nbt-0928	20	10	10	50	408.813	230,241	14085	0	14085	2	14051.20031	2,840,000	4660	563.1939297

⚡ KW17 shows recovery after KW16’s regression, stabilizing win rates with improved Elo. It achieves parity with the FMGo7 baseline while maintaining good efficiency.

🚀 Usage Instructions

1. Download the model

wget https://github.com/changcheng967/kata_web/releases/download/v1.0/kw17-b28c512nbt-0928.bin.gz
mv kw17-b28c512nbt-0928.bin.gz ~/KataGo/models/

2. Run with GTP

./cpp/main gtp \
  -model ~/KataGo/models/kw17-b28c512nbt-0928.bin.gz \
  -config cpp/configs/gtp.cfg

Example session:

boardsize 19
clear_board
genmove B

3. GUI Integration

Sabaki → add engine: ./cpp/main, args: gtp -model models/kw17-b28c512nbt-0928.bin.gz
Lizzie → configure with same GTP arguments
KaTrain → add as custom GTP engine

📦 Files Included

File	Purpose
kw17-b28c512nbt-0928.bin.gz	Final exported model
training_log.txt	Training log and metrics for KW17

Full Changelog: kw16-b28c512nbt-0928...kw17-b28c512nbt-0928

Assets 4

29 Sep 13:30

changcheng967

kw16-b28c512nbt-0928

54bb708

kw16-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw16-b28c512nbt-0928

🔗 Trained using the KataGo open-source framework — full credit to KataGo authors, license, and required attributions.

📌 Overview

kw16 is a fine-tuned 19x19 Go model based on the KW15 checkpoint (kw15-b28c512nbt-0928).
It was trained for 1 epoch on 700,000 samples, totaling 2,640,000 training steps. This continues the KW-series experiments, with improved efficiency and stronger performance compared to KW14.

🧠 Model Information

Attribute	Value
Model Name	kw16-b28c512nbt-0928.bin.gz
Configuration	b28c512nbt (28 blocks, 512 channels)
Board Size	19x19
File Size	~330 MB
Base Model	kw15-b28c512nbt-0928
Training Steps	2,640,000
Training Data	700k rows × 1 epoch of shuffled self-play
Training Time	~1 hour (A100-class GPU, approximate)
Framework	KataGo v1.17.0+ (PyTorch export)

📊 Performance Metrics

Model	Games	Wins	Loses	Win rate (%)	Avg Time (s)	Total Playouts	Baseline Elo	Elo Diff	Calculated Elo	Rank	Weighted Elo	Samples	Elo Realistic
Baseline (FMGo7)		0	14085	1	#VALUE!		14085						4660
KW10-b28c512nbt-0926	20	6	14	30	341.182	192,826	14085	-155.63	13929.37	6	10533.09615	50,000	4643
KW11-b28c512nbt-0926	20	9	11	45	351.037	195,286	14085	-35.22	14049.78	2	10606.35553	200,000	4656
KW12-b28c512nbt-0926	20	8	12	40	369.678	207,036	14085	-70.44	14014.56	4	10539.21332	400,000	4652
KW13-b28c512nbt-0926	20	5	15	25	369.541	210,117	14085	-226.07	13858.93	7	10418.51515	600,000	4635
KW14-b28c512nbt-0927	20	7	13	35	361.964	202,978	14085	-120.41	13964.59	5	10516.77178	1,240,000	4647
KW15-b28c512nbt-0928	20	9	11	45	333.465	186,586	14085	-38.38	14046.62	3	10642.93645	1,940,000	4656
KW16-b28c512nbt-0928	20	7	13	35	364.026	207,507	14085	-155.63	13929.37	N/A	10481.71731	2,640,000	4643

⚡ KW16 did not improve at all(Only Stabality), instead it dropped around 120 ELO and became one of the wordt performanced model, which is one it's not the latest release

🚀 Usage Instructions

1. Download the model

wget https://github.com/changcheng967/kata_web/releases/download/kw16-b28c512nbt-0928/kw16-b28c512nbt-0928.bin.gz
mv kw16-b28c512nbt-0928.bin.gz ~/KataGo/models/

2. Run with GTP

./cpp/main gtp \
  -model ~/KataGo/models/kw16-b28c512nbt-0928.bin.gz \
  -config cpp/configs/gtp.cfg

Example session:

boardsize 19
clear_board
genmove B

3. GUI Integration

Sabaki → add engine: ./cpp/main, args: gtp -model models/kw16-b28c512nbt-0928.bin.gz
Lizzie → configure with same GTP arguments
KaTrain → add as custom GTP engine

📦 Files Included

File Purpose

kw16-b28c512nbt-0928.bin.gz Final exported model

training_log.txt

File	Purpose
kw16-b28c512nbt-0928.bin.gz	Final exported model
training_log.txt	What's Changed Pull commits from main by @changcheng967 in #11 Change readme and remove unneccesary files by @DouletMedia2025 in #17 Training by @changcheng967 in #20 Revert "Training" by @changcheng967 in #21 New Contributors @DouletMedia2025 made their first contribution in #17 Full Changelog: `kw15-b28c512nbt-0928...kw16-b28c512nbt-0928`

What's Changed

Pull commits from main by @changcheng967 in #11
Change readme and remove unneccesary files by @DouletMedia2025 in #17
Training by @changcheng967 in #20
Revert "Training" by @changcheng967 in #21

New Contributors

@DouletMedia2025 made their first contribution in #17

Full Changelog: kw15-b28c512nbt-0928...kw16-b28c512nbt-0928

Contributors

changcheng967 and DouletMedia2025

Assets 4

28 Sep 21:58

changcheng967

kw15-b28c512nbt-0928

e1ce029

kw15-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw15-b28c512nbt-0928

🔗 Trained using the KataGo open-source framework — full credit to KataGo authors, license, and required attributions.

📌 Overview

kw15 is a fine-tuned 19x19 Go model based on the KW14 checkpoint (kw14-b28c512nbt-0927).
It was trained for 1 epoch on 700,000 samples, totaling 1,940,000 training steps. This continues the KW-series experiments, with improved efficiency and stronger performance compared to KW14.

🧠 Model Information

Attribute	Value
Model Name	kw15-b28c512nbt-0928.bin.gz
Configuration	b28c512nbt (28 blocks, 512 channels)
Board Size	19x19
File Size	~330 MB
Base Model	kw14-b28c512nbt-0927
Training Steps	1,940,000
Training Data	700k rows × 1 epoch of shuffled self-play
Training Time	~1 hour (A100-class GPU, approximate)
Framework	KataGo v1.17.0+ (PyTorch export)

📊 Performance Metrics

Model	Games	Wins	Loses	Win rate (%)	Avg Time (s)	Total Playouts	Baseline Elo	Elo Diff	Calculated Elo	Rank	Weighted Elo	Samples	Elo Realistic
Baseline (FMGo7)		0	14085	1	#VALUE!		14085						4660
KW10-b28c512nbt-0926	20	6	14	30	341.182	192,826	14085	-155.63	13929.37	6	10533.09615	50,000	4643
KW11-b28c512nbt-0926	20	9	11	45	351.037	195,286	14085	-35.22	14049.78	2	10606.35553	200,000	4656
KW12-b28c512nbt-0926	20	8	12	40	369.678	207,036	14085	-70.44	14014.56	4	10539.21332	400,000	4652
KW13-b28c512nbt-0926	20	5	15	25	369.541	210,117	14085	-226.07	13858.93	7	10418.51515	600,000	4635
KW14-b28c512nbt-0927	20	7	13	35	361.964	202,978	14085	-120.41	13964.59	5	10516.77178	1,240,000	4647
KW15-b28c512nbt-0928	20	9	11	45	333.465	186,586	14085	-38.38	14046.62	3	10642.93645	1,940,000	4656
KW16-b28c512nbt-0928	20	0	20	0			14085		#VALUE!			2,640,000	#VALUE!

⚡ KW15 shows a clear improvement over KW14 and KW11, with higher efficiency (lower average time, fewer playouts) and stronger Weighted Elo performance.

🚀 Usage Instructions

1. Download the model

wget https://github.com/changcheng967/kata_web/releases/download/v1.0/kw15-b28c512nbt-0928.bin.gz
mv kw15-b28c512nbt-0928.bin.gz ~/KataGo/models/

2. Run with GTP

./cpp/main gtp \
  -model ~/KataGo/models/kw15-b28c512nbt-0928.bin.gz \
  -config cpp/configs/gtp.cfg

Example session:

boardsize 19
clear_board
genmove B

3. GUI Integration

Sabaki → add engine: ./cpp/main, args: gtp -model models/kw15-b28c512nbt-0928.bin.gz
Lizzie → configure with same GTP arguments
KaTrain → add as custom GTP engine

📦 Files Included

File	Purpose
kw15-b28c512nbt-0928.bin.gz	Final exported model
training_log.txt

Assets 4

27 Sep 19:56

changcheng967

kw14-b28c512nbt-0927

9225d59

kw14-b28c512nbt-0927

🏆 kata_web Custom Fine-Tuned Model Release: kw14-b28c512nbt-0927

🔗 Trained using the KataGo open-source framework — full credit to KataGo authors, license, and required attributions.

📌 Overview

kw14 is a fine-tuned 19x19 Go model based on the official KataGo foundation checkpoint (kata1-b28c512nbt-s10904468224-d5317014586).
It was trained for 2 epochs on 620,000 samples per epoch, totaling 1,240,000 samples, continuing the KW-series experiments with deeper data exposure.

🧠 Model Information

Attribute	Value
Model Name	kw14-b28c512nbt-0927.bin.gz
Configuration	b28c512nbt (28 blocks, 512 channels)
Board Size	19x19
File Size	~330 MB
Base Model	kata1-b28c512nbt-s10904468224-d5317014586
Training Steps	~1,240,000 samples (2 epochs)
Training Data	620k rows × 2 epochs of shuffled self-play
Training Time	~1 hour (A100-class GPU, approximate)
Framework	KataGo v1.17.0+ (PyTorch export)

📊 Performance Metrics

Final Loss: reported from training only
Validation: none used (self-play fine-tune)
SWA: ❌ disabled

⚠️ Strength improvement over KW13 is modest; Elo trails KW11 and KW12 in match testing.

⚙️ Training & Export

Training Command

python train.py \
  -traindir /notebooks/katago/shuffledata \
  -datadir /notebooks/katago/shuffledata \
  -exportdir /notebooks/katago/models \
  -exportprefix "kw14-b28c512nbt-0927" \
  -initial-checkpoint /notebooks/katago/initial_checkpoint/kata1-b28c512nbt-s10904468224-d5317014586/model.ckpt \
  -use-fp16 \
  -batch-size 64 \
  -pos-len 19 \
  -max-epochs-this-instance 2 \
  -lr-scale-auto \
  -epochs-per-export 1 \
  -samples-per-epoch 620000

Export Command

python export_model_pytorch.py \
  -checkpoint /notebooks/katago/models/kw14-b28c512nbt-0927-s10905030448-d612345/model.ckpt \
  -export-dir /notebooks/katago/models \
  -model-name kw14-b28c512nbt-0927 \
  -filename-prefix kw14-b28c512nbt-0927

🚀 Usage Instructions

1. Download the model

wget https://github.com/changcheng967/kata_web/releases/download/v1.0/kw14-b28c512nbt-0927.bin.gz
mv kw14-b28c512nbt-0927.bin.gz ~/KataGo/models/

2. Run with GTP

./cpp/main gtp \
  -model ~/KataGo/models/kw14-b28c512nbt-0927.bin.gz \
  -config cpp/configs/gtp.cfg

Example session:

boardsize 19
clear_board
genmove B

3. GUI Integration

Sabaki → add engine: ./cpp/main, args: gtp -model models/kw14-b28c512nbt-0927.bin.gz
Lizzie → configure with same GTP arguments
KaTrain → add as custom GTP engine

📦 Files Included

File	Purpose
kw14-b28c512nbt-0927.bin.gz	Final exported model
training_log.txt	Training log (1.24M samples, 2 epochs)
README.md	This release note

🌟 Highlights

Fine-tuned with 1.24 million samples over 2 epochs
Fully reproducible with clear commands
Useful for education, experiments, and pipeline testing
Performance slightly below KW11/KW12, but stronger than KW13

📝 License

This release complies with the KataGo license.
Full credit to the KataGo team for their work and open-source framework.

What's Changed

Potential fix for code scanning alert no. 2: Workflow does not contain permissions by @changcheng967 in PR #16
Potential fix for code scanning alert no. 1: Workflow does not contain permissions by @changcheng967 in PR #15

Full Changelog: KW-20250919-b18c384nbt-71k-9x9-final → kw14-b28c512nbt-0927

Assets 4

Releases: changcheng967/Kata_web

KW29-b18c384nbt-1018

🏆 KW29 Custom Fine‑Tuned Model Release: kw29‑b18c384nbt‑1018

📦 Overview

🧪 Experiment Setup

📊 Results Summary

Match Log Metrics

Elo & Win Rates

Derived Metrics

📈 Key Insights

✅ Next Steps

🙌 Credits

Uh oh!

KW28-b18c384nbt-1017

Uh oh!

KW27-b18c384nbt-1017

Uh oh!

KW20-b28c512nbt-0930

Uh oh!

KW19-b28c512nbt-0930.bin

What's Changed

Contributors

Uh oh!

kw18-b28c512nbt-0929

🏆 kata_web Custom Fine-Tuned Model Release: kw18-b28c512nbt-0929

📌 Overview

🧠 Model Information

📊 Performance Metrics

🚀 Usage Instructions

1. Download the model

2. Download checkpoint (optional)

3. Run with GTP

Example session:

4. GUI Integration

📦 Files Included

Uh oh!

kw17-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw17-b28c512nbt-0928

📌 Overview

🧠 Model Information

📊 Performance Metrics

🚀 Usage Instructions

1. Download the model

2. Run with GTP

Example session:

3. GUI Integration

📦 Files Included

Uh oh!

kw16-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw16-b28c512nbt-0928

📌 Overview

🧠 Model Information

📊 Performance Metrics

🚀 Usage Instructions

1. Download the model

2. Run with GTP

Example session:

3. GUI Integration

📦 Files Included

What's Changed

New Contributors

Contributors

Uh oh!

kw15-b28c512nbt-0928

🏆 kata_web Custom Fine-Tuned Model Release: kw15-b28c512nbt-0928

📌 Overview

🧠 Model Information

📊 Performance Metrics

🚀 Usage Instructions

1. Download the model

2. Run with GTP

Example session:

3. GUI Integration

📦 Files Included

Uh oh!

kw14-b28c512nbt-0927

🏆 kata_web Custom Fine-Tuned Model Release: kw14-b28c512nbt-0927

📌 Overview

🧠 Model Information

📊 Performance Metrics