HIKARI-Subaru-8B-SkinGroup

Healthcare-oriented Intelligent Knowledge Augmented Retrieval and Inference
Named after Subaru (昴) — the Pleiades star cluster, a group of stars for a model that classifies skin disease groups

---

📦 Model Type: Merged Full Model

This is a fully merged model — the LoRA adapter weights have been merged directly into the base model weights.

✅ No adapter loading needed. Load and run directly with transformers, vLLM, or SGLang.

💾 Size: ~17 GB (4 safetensor shards)

---

Overview

HIKARI-Subaru is Stage 1 of the HIKARI dermatology pipeline — a 4-class skin disease group classifier that divides skin lesions into broad disease families. Its output feeds the Stage 2 disease classifier as context, enabling more precise 10-class diagnosis.

Property	Value
Task	4-class skin disease group classification (Stage 1)
Base model	Qwen/Qwen3-VL-8B-Thinking
Training	Unsloth + LoRA (Fuzzy Top-K sampling)
Val accuracy	88.68% on SkinCAP validation set
Model type	Merged full model

---

🩺 4 Disease Groups

Group	Included Diseases
inflammatory	atopic_dermatitis, psoriasis, seborrheic_dermatitis, urticaria
infectious	acne_vulgaris, tinea_versicolor
neoplastic	melanocytic_nevi, sccis (Bowen's disease), skin_tag
photodermatoses	photodermatoses

---

🔧 Usage

Stage 1 in the Full HIKARI Pipeline

📷 Image
   │
   ▼
[Stage 1] HIKARI-Subaru-8B-SkinGroup ──► group label  ← YOU ARE HERE
   │
   ▼
[Stage 2] HIKARI-Sirius-8B-SkinDx-RAG ──► disease label
   │
   ▼
[Stage 3] HIKARI-Vega-8B-SkinCaption-Fused ──► clinical caption

---

Quick Inference — transformers

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
import torch
from PIL import Image

model_id = "E27085921/HIKARI-Subaru-8B-SkinGroup"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)

image = Image.open("skin_lesion.jpg").convert("RGB")

PROMPT = (
    "Look at this skin lesion image. Based on the visual features, classify it into "
    "one of these groups: inflammatory, infectious, neoplastic, or photodermatoses. "
    "What group does this lesion belong to?"
)

messages = [{"role": "user", "content": [
    {"type": "image", "image": image},
    {"type": "text", "text": PROMPT},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=32, temperature=0.0, do_sample=False)

group = processor.batch_decode(
    out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True
)[0].strip()
print(group)  # → "inflammatory"

---

Production — vLLM BnB-4bit ⚡

from vllm import LLM, SamplingParams
from transformers import AutoProcessor
from PIL import Image

model_id = "E27085921/HIKARI-Subaru-8B-SkinGroup"

llm = LLM(
    model=model_id,
    quantization="bitsandbytes",
    load_format="bitsandbytes",
    trust_remote_code=True,
    max_model_len=2048,
    gpu_memory_utilization=0.88,
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
sp = SamplingParams(max_tokens=32, temperature=0.0)

PROMPT = (
    "Look at this skin lesion image. Based on the visual features, classify it into "
    "one of these groups: inflammatory, infectious, neoplastic, or photodermatoses. "
    "What group does this lesion belong to?"
)

def classify_group(image: Image.Image) -> str:
    messages = [{"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": PROMPT},
    ]}]
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    n = max(text.count("<|vision_start|>"), 1)
    out = llm.generate({"prompt": text, "multi_modal_data": {"image": [image] * n}}, sp)
    return out[0].outputs[0].text.strip()

img = Image.open("skin_lesion.jpg").convert("RGB")
print(classify_group(img))  # → "inflammatory"

---

🌟 HIKARI Model Family

Model	Task	Metric	Type
HIKARI-Subaru-8B-SkinGroup (this model)	4-class group — Stage 1	88.68%	Merged
⭐ HIKARI-Sirius-8B-SkinDx-RAG	10-class disease — Stage 2	85.86%	Merged + LoRA
⭐ HIKARI-Vega-8B-SkinCaption-Fused	Clinical caption — Stage 3	BLEU-4: 29.33	Merged + LoRA

---

📄 Citation

@misc{hikari2026,
  title  = {HIKARI: RAG-in-Training for Skin Disease Diagnosis
            with Cascaded Vision-Language Models},
  author = {Watin Promfiy and Pawitra Boonprasart},
  year   = {2026},
  institution = {King Mongkut's Institute of Technology Ladkrabang,
                 Department of Information Technology, Bangkok, Thailand}
}

---

Made with ❤️ at King Mongkut's Institute of Technology Ladkrabang (KMITL)
Department of Information Technology