HIKARI-Polaris-8B-SkinDx-Oracle

Healthcare-oriented Intelligent Knowledge Augmented Retrieval and Inference
Named after Polaris — the fixed North Star, used as an immovable reference point

---

📦 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 directly with transformers, vLLM, or SGLang.

💾 Size: ~17 GB (4 safetensor shards)

---

⚠️ Research Model — Not for Production

This model requires ground-truth disease group labels at inference time (oracle setting). It is an ablation study model from the HIKARI M-series experiments, designed to measure the theoretical upper bound of the group-context injection approach.

For production use, see:

⭐ E27085921/HIKARI-Sirius-8B-SkinDx-RAG — 85.86% accuracy, no oracle required

---

Overview

HIKARI-Polaris is the M1 GT-Oracle model from our M-series ablation study. It is trained with ground-truth disease group labels injected into both the training prompts and the inference prompts. This tells us: "if we perfectly know the group at inference time, what is the maximum accuracy achievable by the group-injection approach?"

Property	Value
Task	10-class skin disease diagnosis — M1 oracle ablation
Base model	Qwen/Qwen3-VL-8B-Thinking
Context at training	GT group label (oracle)
Context at inference	GT group label (oracle)
Val accuracy (full benchmark)	59.38%
Val accuracy (ablation oracle)	66.0%
Model type	Merged full model

M-Series Ablation Summary

Model	Train Context	Infer Context	Accuracy
M0	None	None	57.4%
M1 / Polaris (this model)	GT group	GT group	66.0% (oracle) / 59.38% (full)
M2	GT group	None	57.4%
M3	Predicted group	Predicted group	61.0%
M4	Predicted group	None	61.0%
HIKARI-Sirius (RAG-in-Training)	RAG references	RAG references	85.86%

Key finding: Even perfect GT-oracle group injection (M1: 66%) is far below RAG-in-Training (Sirius: 85.86%). Group labels alone are insufficient — the model needs visual reference examples to differentiate fine-grained diseases within a group.

---

🔧 Quick Inference — transformers

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

model_id = "E27085921/HIKARI-Polaris-8B-SkinDx-Oracle"

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")

# ⚠️ Requires GROUND-TRUTH group — for research evaluation only
gt_group = "inflammatory"

PROMPT = (
    "This skin lesion belongs to the group '{group}'. "
    "Examine the lesion morphology (papules, plaques, macules), "
    "color (red, violet, white, brown), scale/crust, border sharpness, "
    "and distribution pattern. Based on these visual features, "
    "what is the specific skin disease?"
)

messages = [{"role": "user", "content": [
    {"type": "image", "image": image},
    {"type": "text", "text": PROMPT.format(group=gt_group)},
]}]
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=64, temperature=0.0, do_sample=False)

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

---

📄 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)