Viyavi-Edge-0.8B

Viyavi-Edge-0.8B is a ultra-efficient, fine-tuned Vision-Language Model optimized for Dhivehi OCR in resource-constrained environments. It achieves a sub-1GB RAM footprint while maintaining task-locking on the Thaana script.

Model Details

  • Architecture: Qwen3.5-0.8B-VL-Instruct
  • Task: Dhivehi OCR (Edge/Mobile Deployment)
  • Memory Footprint: 842 MB (4-bit NF4)
  • Language: Dhivehi (dv)
  • Script: Thaana

Performance (Robust Evaluation)

  • Historical Archival: 22.53% CER
  • Modern Student Handwriting: 35.75% CER

Training Configuration

  • Hardware: Single NVIDIA RTX 4090 (24GB VRAM)
  • Framework: Unsloth / PEFT
  • Optimizer: Paged AdamW 8-bit
  • Learning Rate: 2e-4
  • LR Scheduler: Linear with 100-step warmup
  • Effective Batch Size: 8
  • Training Steps: 10,000
  • Quantization: 4-bit NormalFloat (NF4)

Technical Innovations

  • Native VL Integration: Unlike text-only models paired with external encoders, this variant uses a natively integrated visual encoder for maximum efficiency.
  • Resource Optimized: Specifically designed for island council offices and offline digitization tasks on entry-level hardware.

Usage

from peft import PeftModel
from transformers import AutoModelForVision2Seq

base_model = AutoModelForVision2Seq.from_pretrained("Qwen/Qwen3.5-0.8B-VL-Instruct")
model = PeftModel.from_pretrained(base_model, "Ubaidhullah/Viyavi-Edge-0.8B")
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support