uzairkhn/Almas-Pashto-AI

Almas Pashto AI

Almas Pashto AI is a custom, standalone Pashto's first ever Vision-Language Model (VLM) fine-tuned on the 4-Billion parameter google/gemma-3-4b-it base architecture.

It has been meticulously fine-tuned on a huge amount of high-quality Pashto data to deeply understand, generate, and reason in the Pashto language. Crucially, it completely retains its native visual processing capabilities (such as Optical Character Recognition and complex image analysis), making it a fully functioning multimodal assistant natively fluent in Pashto.

Model Details

• Base Model: google/gemma-3-4b-it
• Architecture: Standalone (Weights fully merged)
• Language(s): Pashto (Primary), English
• Capabilities: Text Generation, Vision-Language (Image Analysis, OCR)

How to Load and Use

Because this model is a fully merged standalone architecture, you can load it directly using standard Hugging Face transformers libraries without needing any separate adapter configurations.

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText, BitsAndBytesConfig

# Direct repository ID
model_id = "uzairkhn/Almas-Pashto-AI"

print("Loading processor...")
processor = AutoProcessor.from_pretrained(model_id)

print("Configuring 4-bit quantization...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

print("Loading Almas Pashto AI...")
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto"
)

# Example Inference
test_prompt = "مصنوعي استخبارات څه شی دی؟"

messages = [
    {"role": "user", "content": [{"type": "text", "text": test_prompt}]}
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        repetition_penalty=1.1
    )

generated_text = processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(generated_text)