Ara-JEV

Ara-JEV

Picks the best answer label from a short Arabic situation, a question and optional choices. Made for developers who route or decide Arabic tasks with a small 0.8B-class model.

Made by AhmedReda25300 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ara-JEV decides between answer options on Arabic text. Its base is Qwen3.5-0.8B, with a LoRA added in training. You give it a short situation, a question and optional choices, and it selects the best answer label. This repository holds a merged checkpoint, so you do not load a separate adapter.

Who it suits: It suits developers who need quick, constrained decisions on Arabic text, such as picking an option or setting a priority.

Industry | Who uses it | How they use it | Example
Customer support | Support tool builders | Rate the priority of a ticket written in Arabic | Choose between low, medium and high for a data-exposure report
Software | App developers | Route a request to the right branch of a workflow | Answer a yes or no question about a described situation
Research | NLP researchers | Study small models on Arabic decision data | Score the model on labelled choice questions
Operations | Workflow engineers | Add a quick check to an Arabic pipeline | Ask whether an event has happened yet

If you need open-ended Arabic writing or general Arabic understanding, you can skip it.

Where to use it: It runs in Python with the transformers library, or as a vLLM server.

• Python scripts using transformers
• A vLLM server with the chat-completions API
• Windows, macOS or Linux machines that can run PyTorch

Use case | Who | What they get back
Yes or no decisions | Developers | The chosen Arabic option text
Choice and score questions | Pipeline builders | The index of the selected option, counted from zero
Priority rating | Support teams | A rating index for a described report

How to set it up: You need Python with pip, PyTorch, and transformers at version 5.17 or later, which Qwen3.5 models require. A GPU helps but is not named as required.

1. Install transformers and torch.
2. Load the model with the code below, which needs trust_remote_code turned on.
3. For serving instead, install vLLM and start the server command shown.

pip install "transformers>=5.17" torch

How to use it: You call the model's predict method with a state, a question and a list of options. Arabic yes or no options return the chosen text, and other lists return the index of the chosen option.

import torch
from transformers import AutoModelForImageTextToText

model_id = "AhmedReda25300/Ara-JEV"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()

result = model.predict(
    state="الاجتماع يبدأ الساعة العاشرة صباحًا. دلوقتي الساعة تسعة.",
    question="هل بدأ الاجتماع؟",
    options=["نعم", "لا"],
)
print(result)

To serve it instead, run vLLM with the chat-completions API and limit the response to the allowed labels.

vllm serve AhmedReda25300/Ara-JEV \
  --dtype bfloat16 \
  --max-model-len 8192

Tool | What it does
predict | Builds the prompt and scores the options for you
vllm serve | Serves the model through a chat-completions API
hf download | Fetches the model files to your computer

Example: You pass the situation "The meeting starts at ten in the morning, and it is nine now", ask whether the meeting has started, and offer the two options yes and no. The model returns the Arabic word for no.

For newcomers and experts: Think of this as a multiple-choice grader for Arabic situations. Start with the yes or no example above and check that the printed answer makes sense.

The model is a LoRA of rank 16 trained for one epoch and merged into the base weights. The reported scores use constrained label scoring over known candidates, so they measure choosing among given labels, not free-form generation. A 100-example vLLM test on an A40 averaged 61.3 ms per request.

Give a state, a question and options
The model scores each allowed label
It picks the highest-scoring label
You get the option text or its index

Good to know: Accuracy on the held-out test set of 10,419 examples differs by question type, and yes or no is the weak spot.

Title: Accuracy by question type on the held-out test set
Choice: 98.36 percent
Score: 95.25 percent
Yes or no: 81.71 percent

• Limits: results on dialects, domains and tasks outside the evaluation have not been established.
• Scores are not open-ended reasoning or general Arabic understanding. Validate on your own data first.
• Remote code: loading requires trust_remote_code, which runs the model repository's own predict method.
• License: Apache 2.0, the same as the base model.
• Cost to run: none named.
• Data it reaches: the text you pass in, handled locally once the files are downloaded.
• Not for: high-impact decisions without expert review.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 853M
• Weights: about 1.7 GB in BF16; about 426 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,316
• Likes: 4
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, arabic, qwen3.5, decision-making, merged-lora, text-generation, conversational

Links

• Model: https://huggingface.co/AhmedReda25300/Ara-JEV
• Files: https://huggingface.co/AhmedReda25300/Ara-JEV/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AhmedReda25300/Ara-JEV

Then follow the usage in its model card. Nothing in this kit runs by itself.