README.md

smollm2-360m-physics-gguf

**Author**: [Akhil Vallala](https://www.linkedin.com/in/akhil-fau)  
**Base Model**: [`akhilfau/fine-tuned-smolLM2-360M-with-on-combined_Instruction_dataset`](https://huggingface.co/akhilfau/fine-tuned-smolLM2-360M-with-on-combined_Instruction_dataset)  
**Architecture**: LLaMA (SmolLM2)  
**Parameter count**: 362M  
**Format**: GGUF (Q4_K_M, Q8_0, FP16)  
**License**: Apache 2.0  
**Model Type**: Instruction-tuned Small Language Model (SLM)  
**Use Case**: Solving physics word problems on mobile devices  

Model Overview

This GGUF model is a quantized version of the **Tiny-Physics** model, based on SmolLM2-360M and fine-tuned for **physics word problem solving** using both real and synthetic datasets. It is designed to deliver accurate, low-latency performance on **mobile and edge devices**.

Datasets Used

- 📘 **camel-ai/physics**: Publicly available dataset with 20,000+ physics QA pairs  
- 📘 **Seed dataset**: Extracted from *1000 Solved Problems in Classical Physics*  
- 🧠 **Synthetic dataset**: 6,279 rigorously validated question-answer pairs generated using a GPT-4o-based multi-agent system

These datasets were formatted for instruction tuning using structured prompt–response pairs.

Training Details

- **Model**: SmolLM2-360M  
- **Fine-tuning**: Instruction fine-tuning with LoRA (Low-Rank Adaptation)  
- **Libraries**: Hugging Face Transformers, TRL, Lighteval  
- **Training Epochs**: 3 (best accuracy observed at 3–5 epochs)  
- **Fine-tuning Objective**: Maximize performance on MMLU College Physics  
- **Best Model Accuracy**: `24.51%` on MMLU College Physics  

Evaluation

**Evaluated with**: [Lighteval](https://github.com/huggingface/lighteval)  
**Benchmark**: [MMLU College Physics](https://huggingface.co/datasets/hendrycks_test)  
**Performance**:
| Dataset                      | Accuracy (SmolLM2-360M-Instruct) |
|-----------------------------|----------------------------------|
| MMLU: College Physics       | 24.51%                           |
| Instruction-Tuned camel-ai  | 25.49%                           |
| Combined Instruction Dataset| 24.51%                           |

GGUF Quantization

Model is provided in multiple quantization formats:
| Format   | Size   | Accuracy Retention | Inference Speed | RAM Usage     | Target Use                         |
|----------|--------|--------------------|------------------|---------------|------------------------------------|
| `Q4_K_M` | ~271MB | ~95–97%            | Fast             | ~600–800MB    | Ideal for mid-range mobile devices |
| `Q8_0`   | ~386MB | ~99%               | Medium           | ~1–1.5GB      | Best for higher-end devices        |
| `FP16`   | ~800MB | 100%               | Slow             | ~2GB+         | Reference use only                 |

How to Use

```bash
# Using llama.cpp
./main -m smollm2-360m-physics-gguf.Q4_K_M.gguf -p "What is the acceleration of a 2kg mass falling from 5 meters?"
```

Or via `llama-cpp-python`:

```python
from llama_cpp import Llama

llm = Llama(model_path="smollm2-360m-physics-gguf.Q4_K_M.gguf")
output = llm("What is the potential energy of a 3kg object at 10 meters?")
```

Intended Use

- 📚 **Physics tutoring apps**
- 📶 **Offline mobile inference**
- 🧑‍🏫 **Educational tools for conceptual reasoning**
- 🔋 **Low-power deployment scenarios**

Limitations

- Not trained on multiple-choice formatted data (MCQ output mismatch possible)
- Topic imbalance in datasets may affect generalization
- Not suitable for non-physics or open-domain tasks

Carbon Footprint

Training and fine-tuning consumed approx. **2.64 kg CO₂e**, equivalent to a ~7-mile car ride. This was achieved using local GPU resources (RTX A5500) and energy-efficient batch tuning with LoRA.



Citation

```bibtex
@misc{vallala2025tinyphysics,
  title={Tiny-Physics: A Compact Large Language Model for Physics Word Problems on Mobile Devices},
  author={Akhil Vallala},
  year={2025},
  howpublished={\url{https://huggingface.co/akhilfau/smollm2-360m-physics-gguf}},
}
```