Instructions to use realml/ocr-gemma-3-4b-it with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use realml/ocr-gemma-3-4b-it with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("google/gemma-3-4b-it")
model = PeftModel.from_pretrained(base_model, "realml/ocr-gemma-3-4b-it")

Transformers

How to use realml/ocr-gemma-3-4b-it with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="realml/ocr-gemma-3-4b-it")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("realml/ocr-gemma-3-4b-it", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use realml/ocr-gemma-3-4b-it with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "realml/ocr-gemma-3-4b-it"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "realml/ocr-gemma-3-4b-it",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/realml/ocr-gemma-3-4b-it

SGLang

How to use realml/ocr-gemma-3-4b-it with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "realml/ocr-gemma-3-4b-it" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "realml/ocr-gemma-3-4b-it",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "realml/ocr-gemma-3-4b-it" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "realml/ocr-gemma-3-4b-it",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use realml/ocr-gemma-3-4b-it with Docker Model Runner:
```
docker model run hf.co/realml/ocr-gemma-3-4b-it
```

ocr-gemma-3-4b-it

This model is a fine-tuned version of google/gemma-3-4b-it on the ocr_finetune_train dataset. It achieves the following results on the evaluation set:

Loss: 0.0547

Arabic Legal Documents OCR (Gemma-3-4B-IT)

Fine-tuned Gemma-3-4B-IT model for OCR and structured information extraction from Arabic legal document images.

Model Details

Base model: google/gemma-3-4b-it
Fine-tuning method: LoRA SFT using LLaMA-Factory
Task: OCR + structured JSON extraction
Language: Arabic
Domain: legal / official documents

Repository Structure

arabic-documents-ocr-v1/: merged inference ready model
checkpoints/: intermediate LoRA checkpoints
data/: training datasets and OCR images
training/: LLaMA-Factory training configs
notebooks/: evaluation notebook
adapter_config.json: LoRA adapter configuration

Intended uses & limitations

The model may hallucinate fields if the document is unclear.
OCR quality depends on scan resolution and image preprocessing.
Outputs should be validated before use in legal or production workflows.
Not suitable as a source of legal advice.

This model is intended for:

Arabic OCR extraction
legal document parsing
structured JSON generation from scanned pages
research and experimentation

Training and evaluation data

The model was trained on Arabic legal document page images paired with JSON outputs containing:

markdown-style page content
structural elements
headers / footers
lists and tables
legal articles
document metadata

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 1
eval_batch_size: 1
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 8
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 0.1
num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.1468	1.0101	500	0.1214
0.0806	2.0202	1000	0.0706
0.0643	3.0	1485	0.0547

Framework versions

PEFT 0.18.1
Transformers 5.6.0
Pytorch 2.8.0+cu128
Datasets 4.0.0
Tokenizers 0.22.2

Dataset: https://laws.moj.gov.sa/ar/legislations-regulations?pageNumber=1&pageSize=9&sortingBy=7

Downloads last month: 9

Model tree for realml/ocr-gemma-3-4b-it

Base model

google/gemma-3-4b-pt

Finetuned

google/gemma-3-4b-it

Adapter

(382)

this model