Instructions to use abdullah693/gemma-3-4b-it-urdu-edu-reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use abdullah693/gemma-3-4b-it-urdu-edu-reasoning with Transformers:

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

pipe = pipeline("text-generation", model="abdullah693/gemma-3-4b-it-urdu-edu-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("abdullah693/gemma-3-4b-it-urdu-edu-reasoning")
model = AutoModelForCausalLM.from_pretrained("abdullah693/gemma-3-4b-it-urdu-edu-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use abdullah693/gemma-3-4b-it-urdu-edu-reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "abdullah693/gemma-3-4b-it-urdu-edu-reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "abdullah693/gemma-3-4b-it-urdu-edu-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/abdullah693/gemma-3-4b-it-urdu-edu-reasoning

SGLang

How to use abdullah693/gemma-3-4b-it-urdu-edu-reasoning 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 "abdullah693/gemma-3-4b-it-urdu-edu-reasoning" \
    --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": "abdullah693/gemma-3-4b-it-urdu-edu-reasoning",
		"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 "abdullah693/gemma-3-4b-it-urdu-edu-reasoning" \
        --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": "abdullah693/gemma-3-4b-it-urdu-edu-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use abdullah693/gemma-3-4b-it-urdu-edu-reasoning with Docker Model Runner:
```
docker model run hf.co/abdullah693/gemma-3-4b-it-urdu-edu-reasoning
```

Gemma-3-4B Urdu Education & Reasoning (full fine-tune)

A full fine-tune of google/gemma-3-4b-it for Urdu education, reasoning, and culture, trained on the UrduMMLU-aligned abdullah693/adaption-urdu-edu-cultural-reasoning dataset. The goal is a strong sub-5B Urdu model: math/science reasoning, social-science and Islamic knowledge, Pakistan studies, and native Urdu literature/grammar.

⚠️ Research / educational use. Not an authoritative source for exam preparation, religious rulings, or legal/medical advice.

At a glance


Base	google/gemma-3-4b-it (`Gemma3ForCausalLM`, text)
Method	Full fine-tune (SFT) — all weights updated, not LoRA
Training data	adaption-urdu-edu-cultural-reasoning (~39.9K Urdu/English QA, UrduMMLU-aligned)
Epochs	3 (75 steps)
Loss	2.71 → 0.96 train; eval loss ~1.13
Reported win rate	62% vs. base (pairwise preference)
Languages	Urdu (primary) + English
Trained via	Adaption AutoScientist (Together.AI backend)

Intended use

Urdu question answering, explanation, and reasoning across STEM, humanities/social science, Islamic studies, Pakistan studies, and Urdu language/literature. Prompt it in Urdu; it can produce concise answers or worked explanations, and can be steered into MCQ-style answers.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

mid = "abdullah693/gemma-3-4b-it-urdu-edu-reasoning"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "سورج اور زمین کے درمیان فاصلے کو کیا کہتے ہیں؟ مختصر وضاحت کریں۔"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Evaluation

Designed to be evaluated on UrduMMLU (26,431 Urdu MCQs, 5 domains) using the paper's protocol — generation + parse the option letter, Urdu prompt, 0-/5-shot. Reference baselines (Urdu, 0-shot) for the sub-5B tier:

Model	UrduMMLU (Urdu, 0-shot)
google/gemma-3-4b-it (this model's base)	44.88%
Qwen3-4B	51.70%
LLaMA-3.1-8B	43.84%

Goal: beat the base (~45%) and approach the sub-5B leader. UrduMMLU scores for this fine-tune will be added once evaluated; the training set was deduplicated against UrduMMLU.

Training data composition

The dataset is composed to mirror UrduMMLU's domain mix (rather than over-weighting math): Urdu language/literature/grammar, MMLU humanities & social science (adapted to Urdu), STEM reasoning (GSM8K/MATH/ARC/AQuA), Islamic studies, Pakistan studies, and general knowledge. See the dataset card for full provenance and licenses.

Limitations

4B scale — limited factual depth; can hallucinate, especially on long-tail Pakistan-specific facts.
Urdu grammar and Pakistan geography/current-affairs coverage in training was thin (no permissive native datasets), so these remain weaker areas.
Inherits the biases and knowledge cutoff of the Gemma-3 base.

License

Governed by the Gemma Terms of Use. Fine-tuned by abdullah693; Urdu data adaptation via Adaption AutoScientist.

Citation

@misc{gemma3_4b_urdu_edu_2026,
  title  = {Gemma-3-4B Urdu Education & Reasoning},
  author = {abdullah693},
  year   = {2026},
  url    = {https://huggingface.co/abdullah693/gemma-3-4b-it-urdu-edu-reasoning}
}

Downloads last month: 11

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for abdullah693/gemma-3-4b-it-urdu-edu-reasoning

Base model

google/gemma-3-4b-pt

Finetuned

google/gemma-3-4b-it