s119808-d120396
🏆 KataGo Custom Model Release: mymodel-s119808-d120396
📌 Overview
This is a custom-trained 19x19 Go AI model based on the KataGo framework, trained using a custom training pipeline with a medium-sized dataset. The model has reached amateur mid-low dan level, capable of completing full games while understanding basic Go rules.
🧠 Model Information
| Attribute | Value |
|---|---|
| Model Name | mymodel-s119808-d120396 |
| Model Configuration | b20c256x2 (20 blocks, 256 channels) |
| Board Size | 19x19 |
| File Size | ~83 MB (compressed) |
| Training Data | 120,396 rows (approximately 670 games of 19x19) |
| Training Steps | 119,808 |
| Training Time | ~4 hours (single GPU) |
| Training Framework | KataGo v1.160+ |
📊 Performance Metrics
Training Results
- Training Loss: 40.01
- First Move Accuracy: 28.8%
- Value Variance: 0.076
- Policy Entropy: 0.630
Evaluation Results
| Metric | Value | Description | |——–|——-|————-| | Strength Level | Amateur 5-10 kyu | Can complete full games, understands basic rules | | Eye Formation | Good | Can recognize basic living groups | | Life & Death | Basic | Can handle simple life and death problems | | Endgame | Limited | Will make noticeable mistakes in mid-to-late game | | Stability | Medium | Win rate curve shows moderate fluctuations |
⚙️ Training Parameters
./selfplay/train.sh ~/KataGo/ mymodel b20c256 128 main \
-lr-scale 1.0 \
-max-train-bucket-per-new-data 1 \
-max-train-bucket-size 450000 \
-samples-per-epoch 400000 \
-max-epochs-this-instance 20 \
-sub-epochs 1 \
-max-train-steps-since-last-reload 10000 \
-pos-len 19
🆚 Comparison with Official Models
| Model | Training Samples | First Move Accuracy | Strength Level | File Size |
|---|---|---|---|---|
| This Model | 120k | 28.8% | Amateur 5-10 kyu | ~83 MB |
| g170-b20c256x2 | 668 million | ~45% | Amateur high dan | ~83 MB |
| g170-b40c256x2 | 4.83 billion | ~55%+ | Professional | ~331 MB |
💡 Note: This model was trained with a very small dataset but has reached a playable level. The strongest official model used 4.83 billion samples - 4,000 times more data than this model.
🚀 Usage Instructions
1. Download the model
wget https://github.com/changcheng967/Kata_web/releases/download/v1.2/model.bin.gz
2. Use with KataGo engine
# In KataGo directory
./cpp/katago gtp -model ./models/model.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/model.bin.gz - Lizzie: Configure model path in "Strong Engine Settings"
📦 File Description
model.bin.gz: Compressed model file, ready for use with KataGotrain.log: Complete training log (optional)
🌟 Features & Advantages
- Lightweight: Only requires 83MB storage space
- Quick deployment: Suitable for resource-constrained environments
- Full 19x19 support: Works with standard board without modifications
- Transparent training: Complete training parameters and logs provided
- Beginner-friendly: Ideal for Go AI beginners to study
📅 Future Plans
- Add more training data (target: 50,000+ games)
- Try fine-tuning with official pre-trained models
- Optimize training parameters for better efficiency
- Add 9x9 version model
📝 License
This model follows the KataGo license requirements.
💡 Tip: This model is suitable for entry-level Go AI experience and educational purposes. For higher-level AI, consider using officially released pre-trained models.
Trained based on KataGo open-source framework - Built for Go AI research and education