Forge Coder v1.21.11
A specialized code generation model fine-tuned for Minecraft Forge mod development.
Model Details
| Property | Value |
|---|---|
| Base Model | deepseek-ai/deepseek-coder-6.7b-instruct |
| Fine-tuning Method | LoRA (r=64, alpha=128) |
| Trainable Parameters | 159.9M (2.3% of total) |
| Forge Version | 1.21.11 |
| Minecraft Version | 1.21.11 |
| MCP Version | 20251209.095502 |
Training Data
- Source Code: 27 popular Forge mod repositories
- Documentation: Official Forge documentation
- Total Java Files: 22,916
- Training Samples: 13,936
- Validation Samples: 734
Included Mods
Applied Energistics 2, JustEnoughItems, TerraFirmaCraft, Mekanism, Create, Thermal Expansion/Foundation, RFTools, Botania, Quark, Tinkers' Construct, Immersive Engineering, Twilight Forest, and more.
Training Metrics
| Metric | Value |
|---|---|
| Training Time | 9h 12m |
| Final Train Loss | 0.27 |
| Final Eval Loss | 0.325 |
| Token Accuracy | 92.5% |
| Epochs | 3 |
Capabilities
The model is specialized in:
- Block and Item creation
- Entity programming
- GUI/Screen development
- Network packet handling
- World generation
- Event handling
- Registry systems
- Capability API
- Recipe systems
- Rendering code
- Data generation
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = "deepseek-ai/deepseek-coder-6.7b-instruct"
adapter_path = "path/to/forge-coder-v1.21.11"
tokenizer = AutoTokenizer.from_pretrained(adapter_path)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, adapter_path)
prompt = """### System:
You are an expert Minecraft Forge mod developer.
### User:
Write a simple custom block class for Minecraft Forge 1.21.11
### Assistant:
"""
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))
Version History
- v1.21.11 (2024-12-18): Initial release for MC 1.21.11 / Forge 1.21.11
License
This model is released under the same license as the base model (deepseek-coder). Training data sourced from open-source repositories under various permissive licenses.