---

license: mit datasets:

• phiyodr/coco2017 language:
• en metrics:
• accuracy base_model:
• Qwen/Qwen2-0.5B-Instruct
• google/siglip-base-patch16-224 library_name: transformers pipeline_tag: image-text-to-text

---

Qwiglip VLM (Qwen2 + SigLIP)

Custom Vision-Language Model built from scratch. Inspired by LLaVA VLM architecture, but with a custom MLP projector and LoRA fine-tuning for efficient training. Training data from https://huggingface.co/datasets/phiyodr/coco2017 Full repository at https://github.com/teohyc/qwiglip_vlm

Components

• Base LLM: Qwen/Qwen2-0.5B-Instruct
• Vision Encoder: SigLIP
• LoRA fine-tuning
• Custom MLP projector

Usage

***** CHECK OUT inference.py FOR DETAILED INFERENCE EXAMPLE *****

import torch
from PIL import Image
from transformers import AutoTokenizer, AutoProcessor, AutoModel, Qwen2ForCausalLM
from peft import PeftModel

from vlm_model import MLPProjector, SiglipQwenVLM

#configurations
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

LLM_NAME = "Qwen/Qwen2-0.5B-Instruct"
VISION_NAME = "google/siglip-base-patch16-224"

LORA_PATH = "lora_adapter"
PROJECTOR_PATH = "projector.pt"

NUM_IMAGE_TOKENS = 196

#refer to inference.py for full code