| --- |
| base_model: unsloth/phi-4-mini-reasoning-unsloth-bnb-4bit |
| library_name: peft |
| model_name: ChemPhi-Mini |
| license: apache-2.0 |
| pipeline_tag: text-generation |
| tags: |
| - chemistry |
| - educational-ai |
| - lora |
| - qlora |
| - sft |
| - transformers |
| - trl |
| - unsloth |
| - reasoning |
| - peft |
| - local-llm |
| --- |
| |
| # ChemPhi-Mini |
|
|
| ChemPhi-Mini is a lightweight chemistry-focused reasoning model fine-tuned from `unsloth/phi-4-mini-reasoning-unsloth-bnb-4bit` using supervised fine-tuning (SFT). |
|
|
| This project explores efficient domain adaptation for educational AI systems under constrained hardware environments. The goal was to build a compact chemistry tutoring and reasoning assistant capable of running locally with minimal GPU resources while maintaining useful scientific explanation capabilities. |
|
|
| --- |
|
|
| # Project Goals |
|
|
| This project was built to explore: |
|
|
| - Parameter-efficient fine-tuning (PEFT) |
| - Low-resource LLM training workflows |
| - Chemistry-focused educational reasoning |
| - Lightweight local AI systems |
| - Quantized inference and deployment |
| - Linux-based AI experimentation |
|
|
| The model is part of a broader self-hosted AI and systems engineering learning workflow involving: |
| - Linux infrastructure |
| - Local inference pipelines |
| - GPU-constrained experimentation |
| - Open-source AI tooling |
|
|
| --- |
|
|
| # Base Model |
|
|
| Base model used: |
|
|
| `unsloth/phi-4-mini-reasoning-unsloth-bnb-4bit` |
|
|
| Core characteristics: |
| - Phi-4 Mini Reasoning architecture |
| - 4-bit quantized |
| - Optimized using the Unsloth ecosystem |
| - Designed for efficient fine-tuning and inference |
|
|
| --- |
|
|
| # Training Method |
|
|
| This model was fine-tuned using: |
|
|
| - LoRA (Low-Rank Adaptation) |
| - PEFT |
| - TRL SFTTrainer |
| - 4-bit quantization |
| - Supervised Fine-Tuning (SFT) |
|
|
| Training focused on: |
| - Chemistry explanations |
| - Conceptual reasoning |
| - Educational QA |
| - Scientific response formatting |
|
|
| --- |
|
|
| # Hardware & Environment |
|
|
| Training environment: |
| - Google Colab |
| - NVIDIA T4 GPU |
| - CUDA-enabled PyTorch stack |
|
|
| This project specifically explored practical AI development under limited VRAM conditions. |
|
|
| --- |
|
|
| # Tech Stack |
|
|
| - Transformers |
| - TRL |
| - PEFT |
| - Unsloth |
| - PyTorch |
| - Hugging Face ecosystem |
|
|
| Framework versions: |
| - PEFT 0.19.1 |
| - TRL 0.24.0 |
| - Transformers 5.5.0 |
| - PyTorch 2.10.0+cu128 |
| - Datasets 4.3.0 |
| - Tokenizers 0.22.2 |
|
|
| --- |
|
|
| # Example Usage |
|
|
| ```python |
| from transformers import pipeline |
| |
| generator = pipeline( |
| "text-generation", |
| model="rish3on3AI/ChemPhi-Mini", |
| device="cuda" |
| ) |
| |
| messages = [ |
| { |
| "role": "user", |
| "content": "Explain why increasing temperature favors endothermic reactions." |
| } |
| ] |
| |
| output = generator( |
| messages, |
| max_new_tokens=256, |
| return_full_text=False |
| ) |
| |
| print(output[0]["generated_text"]) |
