KW-20250916-001
🏆 KataGo Custom Fine-Tuned Model Release: KW-20250916-001
📌 Overview
This is a high-performance 19x19 Go AI model, fine-tuned from the powerful kata1-b28c512nbt foundation using a sophisticated two-stage training strategy. The model has reached amateur high-dan level strength, capable of completing full games with sophisticated tactical understanding.
🧠 Model Information
| Attribute | Value |
|---|---|
| Model Name | KW-20250916-001-s10784975104-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 + 100,000 fine-tuning steps |
| Training Data | 43,600 rows |
| Training Time | ~1.5 hours (RTX 5080 laptop GPU) |
| Training Framework | KataGo v1.17.0+ |
📊 Performance Metrics
Final Training Results
- Final Loss: 33.093
- First Move Accuracy: 64.93%
- Value Variance: 0.4486
- Policy Entropy: 0.6535
Validation Results
- Validation Loss: 33.12
- Validation Accuracy: 64.87%
- Validation Variance: 0.447
Strength Assessment
| Metric | Value | Description | |——–|——-|————-| | Strength Level | Amateur 7-8 dan | Strong amateur level | | Eye Formation | Excellent | Can recognize complex living groups | | Life & Death | Good | Can handle most common life and death problems | | Endgame | Medium | Some mistakes in late game | | Middle Game | Excellent | Strong tactical calculation | | Opening | Good | Solid understanding of common patterns |
⚙️ Training Methodology
Two-Stage Fine-Tuning Strategy
Stage 1: Foundation Adaptation
TORCH_LOAD_WEIGHTS_ONLY=0 ./selfplay/train.sh ~/KataGo/ KW-20250916-001-phase1 b28c512nbt 16 main \
-initial-checkpoint ~/KataGo/kata1-b28c512nbt-s10784871168-d5287365110/model.ckpt \
-lr-scale 0.05 \
-max-train-bucket-per-new-data 1 \
-max-train-bucket-size 100000 \
-samples-per-epoch 50000 \
-max-epochs-this-instance 1 \
-sub-epochs 1 \
-max-train-steps-since-last-reload 10000 \
-pos-len 19
Stage 2: Precision Optimization
TORCH_LOAD_WEIGHTS_ONLY=0 ./selfplay/train.sh ~/KataGo/ KW-20250916-001 b28c512nbt 16 main \
-initial-checkpoint ~/KataGo/train/KW-20250916-001-phase1/checkpoint.ckpt \
-lr-scale 0.01 \
-max-train-bucket-per-new-data 1 \
-max-train-bucket-size 100000 \
-samples-per-epoch 50000 \
-max-epochs-this-instance 1 \
-sub-epochs 1 \
-max-train-steps-since-last-reload 10000 \
-pos-len 19
🆚 Comparison with Base Model
| Metric | Base Model | KW-20250916-001 | Improvement |
|---|---|---|---|
| First Move Accuracy | 63.64% | 64.93% | +1.29% |
| Value Variance | 0.457 | 0.4486 | -0.0084 |
| Policy Entropy | 0.621 | 0.6535 | +0.0325 |
| Estimated ELO | ~2350 | ~2375 | +25 |
💡 Note: The two-stage fine-tuning strategy allowed the model to retain the knowledge from the base model while adapting to new data patterns. The slight increase in policy entropy indicates the model has become more decisive in its moves.
🚀 Usage Instructions
1. Download the model
wget https://github.com/changcheng967/Kata_web/releases/download/KW-20250916-001/KW-20250916-KW-20250916-001-s10784975104-d43600.bin.gz
2. Use with KataGo engine
# In KataGo directory
./cpp/katago gtp -model ./models/KW-20250916-001-s10784975104-d43600.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-20250916-001-s10784975104-d43600.bin.gz - Lizzie: Configure model path in "Strong Engine Settings"
- KaTrain: Add to engine list with appropriate parameters
📦 File Description
KW-20250916-001-s10784975104-d43600.bin.gz: Compressed model file, ready for use with KataGoREADME.md: This documentation filetraining_log.txt: Complete training log
🌟 Features & Advantages
- High Strength: 64.93% first move accuracy (amateur 7-8 dan level)
- Full 19x19 Support: Works with standard board without modifications
- Two-Stage Fine-Tuning: Optimized for quality decision-making while preserving base knowledge
- Stable Value Estimation: Low value variance (0.4486) for reliable win rate predictions
- Professional Quality: Suitable for serious study and analysis
📅 Future Plans
- Generate additional high-quality self-play data
- Perform additional fine-tuning rounds with varied learning rates
- Explore ensemble techniques with other strong models
- Create specialized models for specific aspects of Go (life & death, fuseki, etc.)
📝 License
This model follows the KataGo license requirements.
💡 Tip: This model is suitable for serious 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