Specifically, it is fine-tuned under unsloth framework. It was trained using Supervised Fine-Tuning (SFT) with the Low-Rank Adaptation (LoRA) technique. We applied LoRA to both the vision encoder and the language decoder, covering attention projections (q_proj, k_proj, v_proj, o_proj) as well as MLP layers (gate_proj, up_proj, down_proj).

Dataset

Images: 5,000 labeled accommodation images
- Train: 4,700
- Evaluation: 300
- Both training and evaluation splits were balanced within 5% across facility classes and informativeness factors.
Task-specific prompts were used for each image, resulting in multiple samples across different tasks:

Task #Samples
Facility (main) 5,000
Facility (sub) 3,552
Planes 2,268
Meaningful elements 3,552
Proportions 2,400
Building exterior 1,116
View type 1,265
Room type 576
Total 19,729
All prompts were constructed without Chain-of-Thought (CoT) reasoning.

Training Detail

Hyperparameters

LoRA Rank (r): 16
LoRA Alpha (lora_alpha): 32
Learning Rate: 2e-5
Epochs: 2
per_device_train_batch_size: 2
gradient_accumulation_steps: 8

Hardware & Software

GPU: 1x NVIDIA RTX 4090
Framework: unsloth
Training Time: Approximately 8 hours

Intended uses

This model is supervised fine-tuned using specific text prompts and accommodation images. It is recommended to use such text prompts for inference.

How to Use

python

from unsloth import FastVisionModel
import requests
from PIL import Image

# Load model and tokenizer
model, tokenizer = FastVisionModel.from_pretrained(
        model_name = "jwlee9941/informativeness",
        load_in_4bit = False, # Set to False for 16bit LoRA
        max_seq_length = 8192,
)
FastVisionModel.for_inference(model) # Enable for inference!
model.eval()

# Preparation for inference
instruction = "Your task is to identify the type of view visible from the room's interior based on the following criteria:..." # Text Prompt
url = "https://..." # Image URL
image = Image.open(requests.get(url, stream = True).raw)

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": instruction}
    ]}
]
input_text = tokenizer.apply_chat_template(messages, add_generation_prompt = True)
inputs = tokenizer(
    image,
    input_text,
    add_special_tokens = False,
    return_tensors = "pt",
    max_length = 8192,
    truncation = False
).to("cuda")

# Inference: Generation
outputs = model.generate(**inputs, max_new_tokens = 128, do_sample = False)
decoded = tokenizer.decode(outputs[0], skip_special_tokens = True)
print(decoded)

Performance

Metric / Task	Gemini 2.5 Pro	GPT-5	GPT-4o-mini	GLM-4.1V-9B-Thinking	Qwen2.5-VL-7B LoRA (w/o CoT)
Facility (Main)	90.67	92.33	92.33	93.67	92.00
Main&Sub	75.00	82.55	84.91	79.25	85.37
Planes	11.51	46.76	97.12	89.21	97.12
Meaningful Elements	9.43	18.87	38.21	35.85	44.34
Scenery	46.81	31.91	56.03	57.45	67.37
Building Faces	7.35	20.59	8.82	16.18	32.35
Room Type	80.65	83.87	70.97	61.29	87.10
View Type	50.00	64.10	79.49	57.69	74.36

Limitations

Base model constraints: The overall performance is bounded by the capabilities of the chosen base VLM. Using a stronger base model may lead to further improvements.
Single-turn training only: The model was fine-tuned on independent instruction–response pairs rather than multi-turn dialogue data. As a result, it has limited ability to maintain context across multi-turn conversations.
SFT-only stage: Only Stage 1 supervised fine-tuning (SFT) has been applied. Preference alignment methods such as DPO or GRPO were not used, which may limit response stability, consistency, and alignment with user intent. Applying such preference alignment methods as Stage 2 fine-tuning may lead to further improvements.

Task	#Samples
Facility (main)	5,000
Facility (sub)	3,552
Planes	2,268
Meaningful elements	3,552
Proportions	2,400
Building exterior	1,116
View type	1,265
Room type	576
Total	19,729