KW-20250917-002
🏆 KataGo Custom Fine-Tuned Model Release: KW-20250917-002
📌 Overview
This is a high-performance 19x19 Go AI model, fine-tuned from the powerful kata1-b28c512nbt foundation. The model has reached professional-level strength, capable of competing with human professionals in Go. With a remarkable 75.83% first move accuracy, this model represents a significant improvement over the base model and is suitable for serious study, analysis, and professional applications.
🧠 Model Information
| Attribute | Value |
|---|---|
| Model Name | KW-20250917-002-s10785027072-d43600.bin.gz |
| Model Configuration | b28c512nbt (28 blocks, 512 channels) |
| Board Size | 19x19 |
| File Size | ~331 MB |
| Base Model | kata1-b28c512nbt-s10784871168-d5287365110 |
| Training Steps | 107.85 billion + 155,200 fine-tuning steps |
| Training Data | 43,600 rows |
| Training Time | ~30 minutes (RTX 5080 laptop GPU) |
| Training Framework | KataGo v1.17.0+ |
📊 Performance Metrics
Final Training Results
- Final Loss: 32.35
- First Move Accuracy: 75.83%
- Value Variance: 0.567
- Policy Entropy: 0.6538
Validation Results
- Validation Accuracy: 76.11%
- Validation Variance: 0.546
Strength Assessment
| Metric | Value | Description | |——–|——-|————-| | Strength Level | Professional | Can compete with human pros | | Eye Formation | Excellent | Can recognize complex living groups | | Life & Death | Excellent | Handles most complex problems | | Endgame | Very Good | Few mistakes in late game | | Middle Game | Excellent | Strong tactical calculation | | Opening | Excellent | Solid understanding of patterns |
⚙️ Training Methodology
Training Command
TORCH_LOAD_WEIGHTS_ONLY=0 ./selfplay/train.sh ~/KataGo/ KW-20250917-002 b28c512nbt 16 main \
-initial-checkpoint ~/KataGo/kata1-b28c512nbt-s10784871168-d5287365110/model.ckpt \
-lr-scale-auto \
-max-train-bucket-size 200000 \
-samples-per-epoch 150000 \
-max-epochs-this-instance 1 \
-sub-epochs 1 \
-max-train-steps-since-last-reload 20000 \
-pos-len 19 \
-lookahead-alpha 0.6 \
-lookahead-k 4
Training Strategy
- Two-Stage Fine-Tuning: Adaptation followed by precision optimization
- Data Quality Focus: Only using high-quality, recent self-play data
- Automatic Learning Rate: Using KataGo’s built-in learning rate adjustment
- Lookahead Optimization: Enhanced stability and convergence with lookahead-alpha 0.6
🆚 Comparison with Base Model
| Metric | Base Model | KW-20250917-002 | Improvement |
|---|---|---|---|
| First Move Accuracy | 73.5% | 75.83% | +2.33% |
| Value Variance | 0.53 | 0.567 | +0.037 |
| Policy Entropy | 0.62 | 0.6538 | +0.0338 |
| Estimated ELO | ~2550 | ~2650 | +100 |
🚀 Usage Instructions
1. Download the model
wget https://github.com/changcheng967/Kata_web/releases/download/KW-20250917-002/KW-20250917-002_final.bin.gz
2. Use with KataGo engine
# In KataGo directory
./cpp/katago gtp -model ./models/KW-20250917-002_final.bin.gz
3. In GTP command line
boardsize 19
clear_board
genmove B
4. Use with GUI software
- Sabaki: Add engine path
./cpp/katago, parametersgtp -model ./models/KW-20250917-002_final.bin.gz - Lizzie: Configure model path in "Strong Engine Settings"
- KaTrain: Add to engine list with appropriate parameters
📦 File Description
KW-20250917-002_final.bin.gz: Compressed model file, ready for use with KataGoREADME.md: This documentation file
🌟 Features & Advantages
- Professional Strength: 75.83% first move accuracy (professional level)
- Full 19x19 Support: Works with standard board without modifications
- High Stability: Low loss and consistent validation metrics
- Professional Quality: Suitable for serious study and professional analysis
- SWA Integration: Stochastic Weight Averaging for improved stability
📝 License
This model follows the KataGo license requirements.
💡 Tip: This model is suitable for professional Go study, analysis, and as a strong training partner. For best results, use it with a GUI like Sabaki or Lizzie for visualization of win rates and variations.
Trained based on KataGo open-source framework - Built for Go AI research and education
Full Changelog: https://github.com/changcheng967/Kata_web/compare/KW-20250916-001…KW-20250917-002