Mewmew25/Qwen3-4B-Thinking-2507-DocxTrained

Qwen3 4B Thinking - Document Creation Tools

Fine-tuned Qwen3-4B-Thinking-2507 model for creating Word documents (.docx) and Excel files (.xlsx) using JavaScript code generation.

🎯 Capabilities

• ✅ Create Word documents with proper formatting
• ✅ Generate tables and bullet/numbered lists
• ✅ Apply text styling (bold, italic, colors, fonts)
• ✅ Convert Markdown to formatted Word documents
• ✅ Add headers, footers, and page numbers

🚀 Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "your-username/qwen3-4b-docx-tools",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("your-username/qwen3-4b-docx-tools")

# Prepare your prompt
messages = [
    {
        "role": "user",
        "content": "Create a Word document with the title 'Project Report' and three sections with headings. Save it as report.docx"
    }
]

# Generate
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=False)
print(response)

💡 Example Prompts

Simple document:

Create a Word document with the title "Meeting Notes" and save it as meeting.docx

Markdown conversion:

Convert this markdown to Word:

# Main Title
## Section 1
- Point A
- Point B

Save as output.docx

📊 Training Details

• Base Model: Qwen/Qwen3-4B-Thinking-2507
• Training Method: Full model fine-tuning (merged)
• Training Data: 29 curated examples of document creation tasks
• Training Hardware: Kaggle T4 GPU (16GB VRAM)
• Training Time: ~2-3 minutes
• Framework: Unsloth + Hugging Face Transformers

Training Hyperparameters

• Learning Rate: 5e-5
• Epochs: 3
• Batch Size: 2 (effective: 8 with gradient accumulation)
• Max Sequence Length: 1024 tokens
• Optimizer: AdamW 8-bit
• LR Scheduler: Cosine
• Weight Decay: 0.05

📦 Model Size

• Parameters: ~4B
• Model Size: ~8GB (FP16)

🛠️ Requirements

pip install transformers torch docx
# Or for faster inference:
pip install unsloth

⚠️ Limitations

• Trained specifically for document creation tasks
• Requires JavaScript execution environment for generated code
• Limited to 29 training examples (may need more data for edge cases)

📄 License

This model inherits the license from Qwen3-4B-Thinking-2507 (Apache 2.0).

Note: This model generates JavaScript code. Ensure you have a proper execution environment with Node.js and required npm packages installed.