---
language:
- en
license: apache-2.0
library_name: transformers
tags:
- code
- python
- docstring
- documentation
- code-generation
- local-llm
- privacy
- ollama
- qwen3
- knowledge-distillation
- developer-tools
base_model: Qwen/Qwen3-0.6B
pipeline_tag: text-generation
model-index:
- name: Distil-Localdoc-Qwen3-0.6B
results:
- task:
type: text-generation
name: Docstring Generation
metrics:
- type: accuracy
value: 0.76
name: LLM-as-Judge Accuracy
verified: false
---
---
# Distil-Localdoc-Qwen3-0.6B
A small language model (SLM) fine-tuned by Distil Labs for generating high-quality Python docstrings in Google style. Optimized to run locally via Ollama, ensuring your proprietary code never leaves your infrastructure.
*********** [GITHUB DEMO AND CODE](https://github.com/distil-labs/Distil-localdoc/) ***********
## Model Details
- **Developed by**: Distil Labs GmbH
- **License**: Apache 2.0
- **Finetuned from**: Qwen/Qwen3-0.6B
- **Model Size**: 0.6B parameters
- **Deployment**: Local inference via Ollama
## Use-case
Given Python functions or methods without docstrings, the model generates complete, properly formatted documentation following Google style guide.
**Before:**
```python
def calculate_total(items, tax_rate=0.08, discount=None):
subtotal = sum(item['price'] * item['quantity'] for item in items)
if discount:
subtotal *= (1 - discount)
return subtotal * (1 + tax_rate)
```
**After:**
```python
def calculate_total(items, tax_rate=0.08, discount=None):
"""
Calculate the total cost of items, applying a tax rate and optionally a discount.
Args:
items: List of item objects with price and quantity
tax_rate: Tax rate expressed as a decimal (default 0.08)
discount: Discount rate expressed as a decimal; if provided, the subtotal is multiplied by (1 - discount)
Returns:
Total amount after applying the tax
Example:
>>> items = [{'price': 10, 'quantity': 2}, {'price': 5, 'quantity': 1}]
>>> calculate_total(items, tax_rate=0.1, discount=0.05)
22.5
"""
subtotal = sum(item['price'] * item['quantity'] for item in items)
if discount:
subtotal *= (1 - discount)
return subtotal * (1 + tax_rate)
```
The model handles:
- **Functions**: Parameter descriptions, return values, exceptions, and usage examples
- **Methods**: Instance and class method documentation with proper formatting
- **Note**: The tool skips double underscore (dunder: __xxx__) methods
## Why Local?
**Privacy & Security**: Proprietary codebases contain intellectual property and trade secrets. Cloud APIs create:
- IP exposure risks
- Compliance violations (GDPR, SOC 2, HIPAA)
- Security audit failures
- Dependency on external services
**Speed & Cost**: Document entire codebases in minutes without API rate limits or per-token charges.
## Training
The tuned model was trained using knowledge distillation, leveraging the teacher model GPT-OSS-120B. We used 28 diverse Python functions and classes as seed data and supplemented them with 10,000 synthetic examples covering various domains:
- Data science and machine learning
- Web development (Flask, FastAPI, Django)
- DevOps and system utilities
- Algorithm implementations
- API clients and wrappers
Training data includes examples with:
- Various function complexities (simple to async patterns)
- Error handling patterns
- Async/await patterns
- Different parameter types and return values
## Evaluation
We evaluated the model on 250 held-out test examples using LLM-as-a-judge methodology to assess the overall quality of generated docstrings.
| Model | Size | Accuracy |
|--------------------|------|---------------|
| GPT-OSS (thinking) | 120B | 0.81 ± 0.02 |
| Qwen3 0.6B (tuned) | 0.6B | 0.76 ± 0.01 |
| Qwen3 0.6B (base) | 0.6B | 0.55 ± 0.04 |
The fine-tuned model achieves **94%** of the teacher model's performance while running entirely on local hardware with **zero API costs** and **complete privacy**.
## How to Use
### Installation
Follow the instructions in the [Github repository](https://github.com/distil-labs/Distil-localdoc/)
Quick start:
```bash
# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Download and build the model
pip install huggingface_hub
hf download distil-labs/Distil-Localdoc-Qwen3-0.6B --local-dir distil-model
cd distil-model
ollama create localdoc_qwen3 -f Modelfile
# Run on your code
python localdoc_cli.py --file your_script.py
```
### CLI Usage
```bash
# Basic usage (generates Google-style docstrings)
python localdoc_cli.py --file my_module.py
# Use specific model
python localdoc_cli.py --file my_module.py --model localdoc_qwen3
```
The tool will:
1. Parse your Python file using AST
2. Identify all functions and methods without docstrings (skips dunder methods)
3. Generate appropriate docstrings based on code structure
4. Preserve all original code and existing docstrings
5. Output a new file with `_documented` suffix
## Model Sources
- **Homepage**: [https://distillabs.ai](https://distillabs.ai)
- **Repository**: [https://github.com/distil-labs/Distil-localdoc](https://github.com/distil-labs/Distil-localdoc)
- **Contact**: contact@distillabs.ai
## Citation
```bibtex
@software{distil_localdoc_2024,
title = {Distil-Localdoc: Local Python Documentation Generation with SLMs},
author = {Distil Labs},
year = {2024},
url = {https://huggingface.co/distil-labs/Distil-Localdoc-Qwen3-0.6B}
}
```
## Community
- Follow us on [LinkedIn](https://www.linkedin.com/company/distil-labs/)
- Join our [Slack community](https://join.slack.com/t/distil-labs-community/shared_invite/zt-36zqj87le-i3quWUn2bjErRq22xoE58g)
- Star us on [GitHub](https://github.com/distil-labs/Distil-localdoc)
