KW-20250919-b18c384nbt-71k-9x9-final
🏆 KataGo Custom Fine-Tuned Model Release: KW-20250919-b18c384nbt-71k-9x9-final
KataGo
🔗 Trained based on KataGo open-source framework — Built for Go AI research and education
📌 Overview
This is a high-performance 9x9 Go AI model, fine-tuned from the official kata9x9-b18c384nbt-20231025 foundation. After 10 epochs of self-play training on true 9x9 data (board size strictly 9x9), this model achieves strong amateur to low-dan level strength and demonstrates excellent tactical understanding in small-board games. With 80.6% first move accuracy on test positions, it outperforms the base model and is ideal for learning, fast analysis, and AI-assisted 9x9 study.
Perfect for players improving their opening, life-and-death, and middle-game tactics!
🧠 Model Information
| Attribute | Value |
|---|---|
| Model Name | KW-20250919-b18c384nbt-71k-9x9-final-s6604185600-d69367.bin |
| Model Configuration | b18c384nbt (18 blocks, 384 channels) |
| Board Size | 9x9 |
| File Size | ~110 MB |
| Base Model | kata9x9-b18c384nbt-20231025.bin |
| Training Steps | 6.6 billion samples (10 epochs) |
| Training Data | 69,367 rows of clean 9x9 self-play |
| Training Time | ~50 minutes (RTX 5000 class GPU) |
| Training Framework | KataGo v1.17.0+ (PyTorch export) |
📊 Performance Metrics
Final Training Results
- Final Loss: 32.0
- First Move Accuracy (pacc1): 80.6%
- Value Variance (vsquare): 0.59
- Policy Entropy (ptentr): 0.61
- SWA Enabled: ❌False
Validation
- No validation files used (standard for self-play pipelines)
🆚 Comparison with Base Model
| Metric | Base Model (kata9x9-...) |
KW-20250919-9x9-final | Improvement |
|---|---|---|---|
| First Move Accuracy | ~70–72% (estimated) | 80.6% | +~8–10% |
| Training Data Quality | Mixed board sizes | Pure 9x9 only | ✅ Cleaner |
| Board Alignment | Padded 19x19 tensors | True 9x9 layout | ✅ Optimal |
| Estimated Strength | Low Dan (9x9) | High-Dan to Lower Professional(9x9) | ▲ Stronger |
💡 This model learns more accurate local patterns due to consistent 9x9 training.
⚙️ Training Methodology
Training Command
TRAIN_NAME=\"KW-20250919-b18c384nbt-71k-9x9-final\"
TORCH_LOAD_WEIGHTS_ONLY=0 ./selfplay/train.sh \
~/KataGo/ $TRAIN_NAME b18c384nbt 16 main \
-pos-len 9 \
-initial-checkpoint /home/chang/KataGo/models/kata9x9-b18c384nbt-20231025.ckpt \
-lr-scale-auto \
-max-train-bucket-size 100000 \
-samples-per-epoch 56704 \
-max-epochs-this-instance 10 \
-sub-epochs 1 \
>> ~/KataGo/logs/$TRAIN_NAME.log 2>&1
Key Training Features
- ✅ True 9x9 Pipeline: All data generated with
dataBoardLen = 9 - ✅ Stochastic Weight Averaging (SWA): Smoother, more stable weights
- ✅ Lookahead Optimizer: Improves convergence and generalization
- ✅ FP32 Training: Maximum precision
- ✅ Clean Self-Play Data: No human or mixed-size contamination
🚀 Usage Instructions
1. Download the Model
# Example (replace with your actual GitHub release URL)
wget https://github.com/yourusername/KataGo-9x9/releases/download/v1.0/KW-20250919-b18c384nbt-71k-9x9-final.bin
mv KW-20250919-b18c384nbt-71k-9x9-final.bin ~/KataGo/models/
2. Run with GTP (Command Line)
~/KataGo/cpp/main gtp \
-model ~/KataGo/models/KW-20250919-b18c384nbt-71k-9x9-final.bin \
-config cpp/configs/gtp.cfg
Then:
boardsize 9
clear_board
genmove B
3. Use in GUI Software
| Software | Setup Guide |
|———|————-|
| Sabaki | Add engine: ./cpp/main, args: gtp -model models/KW-20250919-b18c384nbt-71k-9x9-final.bin |
| Lizzie / Leela Zero | Load model via "Engine Settings" → point to .bin file |
| KaTrain | Add as custom engine with GTP command above |
📦 File Description
| File | Purpose |
|---|---|
KW-20250919-b18c384nbt-71k-9x9-final.bin |
Final trained 9x9 model (ready to use) |
README.md |
This release note |
training_log.txt |
Full training log (for debugging/analysis) |
🌟 Features & Advantages
- ✅ True 9x9 Architecture: No wasted computation on padding
- ✅ Faster Inference: Ideal for mobile or lightweight devices
- ✅ Excellent for Learning: Great for practicing openings, tactics, and endgames
- ✅ Strong Tactical Play: Excels at life-and-death and local fights
- ✅ Open & Reproducible: Full pipeline documented and shared
📝 License
This model follows the KataGo license requirements.
Trained using open-source data and tools. For research, education, and non-commercial use.
💡 Tip
Use this model in Sabaki or Lizzie to visualize win rate graphs, best variations, and policy heatmaps. It’s an excellent sparring partner for improving your 9x9 game speed and pattern recognition!
KataGo
Trained based on KataGo open-source framework — Built for Go AI research and education
What’s Changed
- Bump actions/configure-pages from 3 to 5 in /.github/workflows by @dependabot[bot] in https://github.com/changcheng967/Kata_web/pull/8
- Bump actions/checkout from 3 to 5 in /.github/workflows by @dependabot[bot] in https://github.com/changcheng967/Kata_web/pull/10
- Bump actions/setup-python from 4 to 6 in /.github/workflows by @dependabot[bot] in https://github.com/changcheng967/Kata_web/pull/9
- Bump actions/upload-pages-artifact from 3 to 4 in /.github/workflows by @dependabot[bot] in https://github.com/changcheng967/Kata_web/pull/7
New Contributors
- @dependabot[bot] made their first contribution in https://github.com/changcheng967/Kata_web/pull/8
Full Changelog: https://github.com/changcheng967/Kata_web/compare/KW-20250917-002…KW-20250919-b18c384nbt-71k-9x9-final