初始化项目，由ModelHub XC社区提供模型

Model: deltakitsune/Nanbeige-4.1-Python-DeepThink-3B Source: Original Platform
2026-08-28 14:08:18 +08:00
commit 9cd3d8d3c0
15 changed files with 815 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,140 @@
+---
+license: apache-2.0
+base_model: Nanbeige/Nanbeige4.1-3B
+tags:
+- code
+- python
+- fine-tuned
+- lora
+- direct-output
+language:
+- en
+pipeline_tag: text-generation
+---
+
+# Nanbeige 4.1 Python DeepThink - 3B
+
+Fine-tuned version of [Nanbeige/Nanbeige4.1-3B](https://huggingface.co/Nanbeige/Nanbeige4.1-3B) specialized for Python code generation with direct, focused output.
+
+**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 <think> 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