Instructions to use michelinolinolino/gemma4-4b-sci with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use michelinolinolino/gemma4-4b-sci with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="michelinolinolino/gemma4-4b-sci",
	filename="model.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use michelinolinolino/gemma4-4b-sci with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf michelinolinolino/gemma4-4b-sci
# Run inference directly in the terminal:
llama-cli -hf michelinolinolino/gemma4-4b-sci

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf michelinolinolino/gemma4-4b-sci
# Run inference directly in the terminal:
llama-cli -hf michelinolinolino/gemma4-4b-sci

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf michelinolinolino/gemma4-4b-sci
# Run inference directly in the terminal:
./llama-cli -hf michelinolinolino/gemma4-4b-sci

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf michelinolinolino/gemma4-4b-sci
# Run inference directly in the terminal:
./build/bin/llama-cli -hf michelinolinolino/gemma4-4b-sci

Use Docker

docker model run hf.co/michelinolinolino/gemma4-4b-sci

LM Studio
Jan

vLLM

How to use michelinolinolino/gemma4-4b-sci with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "michelinolinolino/gemma4-4b-sci"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "michelinolinolino/gemma4-4b-sci",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/michelinolinolino/gemma4-4b-sci

Ollama
How to use michelinolinolino/gemma4-4b-sci with Ollama:
```
ollama run hf.co/michelinolinolino/gemma4-4b-sci
```

Unsloth Studio new

How to use michelinolinolino/gemma4-4b-sci with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for michelinolinolino/gemma4-4b-sci to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for michelinolinolino/gemma4-4b-sci to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for michelinolinolino/gemma4-4b-sci to start chatting

Docker Model Runner
How to use michelinolinolino/gemma4-4b-sci with Docker Model Runner:
```
docker model run hf.co/michelinolinolino/gemma4-4b-sci
```

Lemonade

How to use michelinolinolino/gemma4-4b-sci with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull michelinolinolino/gemma4-4b-sci

Run and chat with the model

lemonade run user.gemma4-4b-sci-{{QUANT_TAG}}

List all available models

lemonade list

gemma4-4b-sci

Early-stage research experiment. Trained for 1 epoch on 30K examples. Expect hallucinations and factual errors.

gemma4-4b-sci is a scientific-domain fine-tune of Gemma 4 E4B via QLoRA on 30,000 examples from OpenSciLM/OS_Train_Data and SciRIFF. Inspired by OpenScholar — this is a generation-only model without a retrieval pipeline.

Model Description

Developed by: Michele Banfi
Base model: unsloth/gemma-4-E4B-it
Method: QLoRA (4-bit) + SFT via Unsloth, language layers only (vision encoder frozen)
Training: 1 epoch, 30K examples (15K OS_Train_Data + 15K SciRIFF), NVIDIA RTX 5090
License: Gemma Terms of Use

Model Sources

Repository: https://github.com/michelebanfi/gemma-4-finetuning
Evaluation: ScholarQABench
Ollama: ollama run hf.co/michelinolinolino/gemma4-4b-sci

Quick Start

ollama run hf.co/michelinolinolino/gemma4-4b-sci

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained("michelinolinolino/gemma4-4b-sci", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("michelinolinolino/gemma4-4b-sci")

messages = [{"role": "user", "content": "Explain the role of CRISPR-Cas9 in gene editing."}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(input_ids, max_new_tokens=512)[0][input_ids.shape[1]:], skip_special_tokens=True))

Evaluation

ScholarQABench — draft results, 1-epoch run. Gold paper contexts provided (fair comparison with OpenScholar-8B).

Task	Metric	gemma4-4b-sci	OpenScholar-8B
SciFact (208)	Accuracy	77.9%	76.4%
PubMedQA (843)	Accuracy	81.5%	76.0%
QASA (1375)	ROUGE-L	20.9	23.0
SciFact	Citation F1	0.0	68.9
PubMedQA	Citation F1	0.0	43.6
QASA	Citation F1	4.3	56.3

Correctness matches or exceeds OpenScholar-8B (2× the parameters) at 1 epoch. Citation gap is entirely due to the missing retrieval pipeline.

Citation

@article{asai2024openscholar,
  title   = {OpenScholar: Synthesizing Scientific Literature with Retrieval-Augmented LMs},
  author  = {Asai, Akari and others},
  journal = {Nature},
  year    = {2024},
  url     = {https://allenai.org/blog/nature-openscilm}
}

Downloads last month: 321

GGUF

Model size

8B params

Architecture

gemma4

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for michelinolinolino/gemma4-4b-sci

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Finetuned

unsloth/gemma-4-E4B-it

Quantized

(11)

this model

michelinolinolino
/

gemma4-4b-sci