Version: E1 (Experiment 1)
Training Focus: Code accuracy and clean output format
Status: Production-ready for direct code generation tasks

Model Description

This model was fine-tuned using LoRA on 45,757 examples (84% Python code, 16% mathematical reasoning) to specialize in Python code generation. It achieves 87.4% token-level accuracy while providing clean, direct responses optimized for production use.

Training Details

Base Model: Nanbeige/Nanbeige4.1-3B (3B parameters)
Method: LoRA (r=16, alpha=16)
Trainable Parameters: 28.4M (0.72%)
Training Time: ~16 hours on RTX 5060 Ti 16GB
Datasets: Magicoder-OSS-Instruct-75K (Python), GSM8K (reasoning)
Framework: Transformers + PEFT

Performance Improvements

Metric	Baseline	Fine-tuned	Change
Loss	1.04	0.45	-57%
Token Accuracy	76.3%	87.4%	+11.1 pts
Entropy	0.78	0.44	-44%

Key Features

✅ Direct Output Format - Clean code responses without verbose preambles
✅ High Accuracy - 87% token-level accuracy on Python tasks
✅ Fast Inference - Optimized for quick responses
⚠️ Suppressed Chain-of-Thought - E1 focuses on direct answers (reasoning occurs internally but isn't narrated)

Usage

Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    'deltakitsune/Nanbeige-4.1-Python-DeepThink-3B',
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    'deltakitsune/Nanbeige-4.1-Python-DeepThink-3B',
    trust_remote_code=True
)

prompt = 'Write a Python function to validate email addresses'
inputs = tokenizer(prompt, return_tensors='pt')
outputs = model.generate(**inputs, max_length=512)
print(tokenizer.decode(outputs[0]))

Ollama

bash

# Pull from Ollama registry
ollama pull fauxpaslife/nanbeige4.1-python-deepthink:3b

# Run
ollama run fauxpaslife/nanbeige4.1-python-deepthink:3b

llama.cpp

bash

# Download GGUF
wget https://huggingface.co/deltakitsune/Nanbeige-4.1-Python-DeepThink-3B/resolve/main/nanbeige4.1-python-deepthink-q8.gguf

# Run
./llama-cli -m nanbeige4.1-python-deepthink-q8.gguf -p \"Write a binary search function\"

File Structure

*.safetensors - Merged model weights (Transformers)
config.json - Model configuration
okenizer.json - Tokenizer files

anbeige4.1-python-deepthink-fp16.gguf - Full precision GGUF (7.9GB)

anbeige4.1-python-deepthink-q8.gguf - 8-bit quantized GGUF (4.2GB)

Best Use Cases

Direct Python code generation
Algorithm implementations
Flask/FastAPI endpoint creation
Code debugging with concise explanations
Production codebases requiring deterministic output

When to Use Base Model Instead

Complex problems requiring visible reasoning
Exploring multiple solution approaches
Educational explanations with thought process
Research/debugging requiring transparency

Training Notes

E1 focused on direct output format. Training data contained no chain-of-thought examples, resulting in suppressed tag behavior. Internal reasoning capability is preserved (evidenced by accuracy gains), but output format is optimized for production code generation.

E2 Development: Next iteration will reintroduce chain-of-thought reasoning while maintaining code quality.

Citation

bibtex

@misc{nanbeige-python-deepthink-e1,
  title={Nanbeige 4.1 Python DeepThink 3B},
  author={deltakitsune},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/deltakitsune/Nanbeige-4.1-Python-DeepThink-3B}
}

License

Apache 2.0 (same as base model)

Developed By

deltakitsune (fauxpaslife)
Part of the Delta:Kitsune AI platform development
February 2026