| # Qwenmark2-0.5B Fine-Tuned Model Overview |
|
|
| This is a fine-tuned version of the Qwen2-0.5B model, a transformer-based language model developed by Alibaba Cloud. The model has been fine-tuned using **LoRA (Low-Rank Adaptation)** and **Unsupervised Parameter-Efficient Fine-Tuning (PFT)** to specialize in deep learning and machine learning educational tasks. |
|
|
| --- |
|
|
| ## ✅ Key Features |
|
|
| - 🎯 **Specialization**: Deep learning & machine learning Q&A |
| - 📘 **Educational Utility**: Enhanced explanation performance |
| - ⚙️ **Efficient Deployment**: Optional 4-bit quantization |
| - 💡 **Contextual Understanding**: Supports RAG-style inference |
|
|
| --- |
|
|
| ## 🧠 Model Details |
|
|
| - **Base Model**: `Qwen/Qwen2-0.5B` |
| - **Architecture**: Transformer-based Causal Language Model |
| - **Parameters**: 0.5 Billion |
| - **Tokenizer**: Qwen2 tokenizer (`left padding`, `eos_token` as `pad_token` if unspecified) |
| - **Quantization**: Supports 4-bit via `BitsAndBytesConfig` |
| - **Devices Supported**: CUDA-enabled GPUs / CPU |
|
|
| --- |
|
|
| ## 🔧 Fine-Tuning Method |
|
|
| ### 1. LoRA Distillation |
| - **Data**: DeepseekR1-generated answers to curated ML/DL questions |
| - **Config**: `r=16`, `lora_alpha=32`, `target_modules=["q_proj", "v_proj"]` |
| - **Training**: 3 epochs, batch size 2, grad_accum=4, lr=2e-4, FP16 |
| - **Output**: `./lora_finetuned` |
|
|
| ### 2. Unsupervised PFT |
| - **Data**: Extracted text from `course_slides_text.txt` |
| - **Training**: 1 epoch, batch size 2, grad_accum=4, lr=1e-5, FP16 |
| - **Output**: `./LoRA&pft_finetuned` |
|
|
| --- |
|
|
| ## 🛠️ Installation |
|
|
| Install required packages: |
| ```bash |
| pip install torch transformers peft datasets sentence-transformers pdf2image pytesseract |
| |
