Gemma 4 Particle Edu — LoRA Adapters (26B + 31B shallow + 31B deep)

GGUF-format LoRA adapters for three Gemma 4 sizes, fine-tuned on 907 Alpaca physics simulation pairs via Unsloth QLoRA.

Companion repository for U2DIA/gemma4-particle-edu-e4b (the smallest fine-tune, shipped as merged Q4_K_M GGUF).

Files

File	Size	Base	LoRA config
`lora-26b-physics.gguf`	36 MB	Gemma 4 26B MoE	QLoRA r=8
`lora-31b-shallow.gguf`	234 MB	Gemma 4 31B Dense	QLoRA r=8, 1 epoch
`lora-31b-deep.gguf`	1.9 GB	Gemma 4 31B Dense	QLoRA r=64, 3 epochs
`Modelfile-26b`	-	-	Ollama Modelfile for 26B
`Modelfile-31b-shallow`	-	-	Ollama Modelfile for 31B shallow
`Modelfile-31b-deep`	-	-	Ollama Modelfile for 31B deep

Training cost

26B: $2.40 on Lambda GH200 96GB
31B shallow (r=8, 1ep): $2.55 on Lambda GH200 96GB
31B deep (r=64, 3ep): $2.55 on Lambda GH200 96GB
Total for the 3 adapters: $7.50
E4B (separate repo): $0.55 on Lambda A10

Benchmark (20 scenarios each)

Model	JSON parse	Physics	Time
Base 9B	30%	0%	12.7s
E4B FT	70%	77%	8.9s
Base 26B MoE	95%	22%	9.3s
26B FT	90%	31%	9.3s
Base 31B	100%	21%	20.6s
31B shallow	100%	18%	21.1s
31B deep	100%	18%	20.0s

Finding: Larger bases (26B/31B) already achieve 95-100% JSON parsing without fine-tuning, so the 907-pair dataset cannot move them further. E4B is the cost-optimal choice.

How to use with Ollama

# Download this repo
huggingface-cli download U2DIA/gemma4-particle-edu-loras --local-dir ./loras

# Pull base model
ollama pull gemma4:31b

# Register fine-tuned variant
cat > Modelfile <<EOF
FROM gemma4:31b
ADAPTER ./loras/lora-31b-deep.gguf
EOF
ollama create gemma4-31b-physics-edu-deep -f Modelfile
ollama run gemma4-31b-physics-edu-deep