U2DIA/gemma4-particle-edu-e4b

Gemma 4 Particle Edu — E4B Fine-tuned (Q4_K_M GGUF)

Fine-tuned Gemma 4 E4B (4.5B active) for physics simulation parameter generation. Part of the Gemma 4 Particle Edu Kaggle Good Hackathon submission.

What this model does

Given a natural language physics scenario (e.g., "DNA double helix at body temperature"), this model outputs a JSON simulation specification with SI-unit physics parameters:

{
  "simulation": {
    "prompt": "dna",
    "title": "DNA Double Helix",
    "domain": "biology",
    "physics": {
      "gravity": 0,
      "damping": 0.99,
      "springStiffness": 30,
      "particleCount": 22000,
      "temperature": 310,
      "density": 1700
    }
  }
}

Training details

• Method: Unsloth QLoRA (r=16)
• Base: Gemma 4 E4B (4.5B active parameters)
• Dataset: 907 Alpaca-format physics simulation pairs
• Hardware: Lambda A10 (24GB)
• Cost: $0.55
• Quantization: llama.cpp Q4_K_M (CPU-only conversion)

Benchmark vs other Gemma 4 sizes

All 4 sizes fine-tuned on the same 907-pair dataset:

Model	Type	JSON parse	Physics	Time	Cost
Base Gemma 4 9B	Dense	30%	0%	12.7s	-
E4B FT (this model)	QLoRA r=16	70%	77%	8.9s	$0.55
Base Gemma 4 26B MoE	MoE	95%	22%	9.3s	-
26B FT	QLoRA r=8	90%	31%	9.3s	$2.40
Base Gemma 4 31B	Dense	100%	21%	20.6s	-
31B shallow FT	r=8, 1ep	100%	18%	21.1s	$2.55
31B deep FT	r=64, 3ep	100%	18%	20.0s	$2.55

Finding: E4B QLoRA is cost-optimal — $0.55 delivers +40%p JSON success and +77%p physics accuracy over the 9B base. Larger bases (26B/31B) already achieve 95-100% JSON parsing, so the 907-pair dataset cannot move them further.

How to use

Ollama (recommended)

# Pull this repo and register with Ollama
huggingface-cli download U2DIA/gemma4-particle-edu-e4b --local-dir ./gemma4-e4b
cd gemma4-e4b
ollama create gemma4-physics-edu -f Modelfile
ollama run gemma4-physics-edu

llama.cpp

./llama-cli -m gemma4-physics-edu-Q4_K_M.gguf -p "Simulate a DNA double helix"

Files

File	Size	Description
gemma4-physics-edu-Q4_K_M.gguf	5.3 GB	Merged Q4_K_M quantized weights
config.json	6 KB	Hugging Face model config
tokenizer.json	31 MB	Tokenizer
Modelfile	241 B	Ollama Modelfile

Related resources

• GitHub: https://github.com/U2SY26/gemma4-particle-edu
• Live Demo: https://gemma4-particle-edu.vercel.app
• Kaggle Writeup: https://www.kaggle.com/competitions/gemma-4-good-hackathon/writeups/gemma-4-particle-edu-free-3d-physics-simulation-v
• Kaggle Benchmark Dataset: https://www.kaggle.com/datasets/syu21125/gemma4-particle-edu-benchmark-300
• Kaggle Ollama Live Demo: https://www.kaggle.com/code/syu21125/gemma-4-particle-edu-ollama-live-demo
• 3dweb (production app, 8,470 installs): https://play.google.com/store/apps/details?id=com.sciencelab.science_lab_flutter

Limitations

• 70% JSON parse rate means ~30% of outputs need retry or fallback
• Physics accuracy was measured on 20 scenarios; full 300-scenario benchmark requires the 31B model
• Fine-tuned on English prompts; Korean prompts fall back to the base model's multilingual capability
• Not suitable for production medical, safety-critical, or regulatory-compliant simulations

Competition

Submitted to Kaggle Gemma 4 Good Hackathon (2026-05-18 deadline).

Tracks: Impact (Education) + Special Technology (Ollama + Unsloth)

Citation

@misc{gemma4-particle-edu-e4b,
  author = {Yun (U2DIA)},
  title = {Gemma 4 Particle Edu — E4B Fine-tuned},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/U2DIA/gemma4-particle-edu-e4b}
}