---
license: apache-2.0
base_model: Qwen/Qwen3-VL-8B-Thinking
tags:
  - dermatology
  - medical
  - vision-language
  - caption-generation
  - clinical-nlp
  - fine-tuned
  - qwen3-vl
language:
  - en
  - th
datasets:
  - SkinCAP
pipeline_tag: image-text-to-text
---

<p align="center">
  <img src="HIKARI_logo.png" alt="HIKARI — Healthcare-oriented Intelligent Knowledge Augmented Retrieval and Inference" width="100%"/>
</p>

<h1 align="center">HIKARI-Rigel-8B-SkinCaption</h1>

<p align="center">
  <b>Healthcare-oriented Intelligent Knowledge Augmented Retrieval and Inference</b><br/>
  Named after <b>Rigel</b> — blue supergiant in Orion, a first step in the caption training path
</p>

<p align="center">
  <img src="https://img.shields.io/badge/Base%20Model-Qwen3--VL--8B--Thinking-blue?style=flat-square"/>
  <img src="https://img.shields.io/badge/Task-Clinical%20Caption%20Generation-teal?style=flat-square"/>
  <img src="https://img.shields.io/badge/BLEU--4-9.82-yellow?style=flat-square"/>
  <img src="https://img.shields.io/badge/Init-Checkpoint--Init-gray?style=flat-square"/>
  <img src="https://img.shields.io/badge/License-Apache%202.0-orange?style=flat-square"/>
</p>

---

## 📦 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)
>
> 🔌 Lightweight adapter version:
> **[E27085921/HIKARI-Rigel-8B-SkinCaption-LoRA](https://huggingface.co/E27085921/HIKARI-Rigel-8B-SkinCaption-LoRA)** (~1.1 GB)

---

## Overview

**HIKARI-Rigel** generates clinical skin lesion captions using the **checkpoint-init** (Way 1) training strategy: Stage 3 caption training continues directly from the Stage 2 LoRA checkpoint, fine-tuning the existing disease adapters further on caption data.

This is an ablation baseline. For best captioning performance, use [⭐ HIKARI-Vega-8B-SkinCaption-Fused](https://huggingface.co/E27085921/HIKARI-Vega-8B-SkinCaption-Fused) (BLEU-4: 29.33, 3× better).

| Property | Value |
|:---------|:------|
| **Task** | Clinical skin lesion caption generation (Stage 3) |
| **Base model** | `Qwen/Qwen3-VL-8B-Thinking` |
| **Init strategy** | Checkpoint-Init — continues from Stage 2 LoRA checkpoint |
| **BLEU-4** | 9.82 |
| **ROUGE-1** | 38.90 |
| **BERTScore-F** | 88.12 (roberta-large) |
| **Model type** | Merged full model |

### Why Checkpoint-Init Underperforms

The Stage 2 disease LoRA adapters are directly continued into caption training. The caption learning signal overwrites the disease knowledge that was stored in those same LoRA weights. Result: the model loses its diagnostic ability before it fully learns to generate captions.

| Init | BLEU-4 | ROUGE-1 | Disease knowledge |
|:-----|:------:|:-------:|:-----------------:|
| **Checkpoint (this model)** | 9.82 | 38.90 | ❌ Lost during training |
| **Merged (Vega)** | **29.33** | **53.55** | ✅ Locked in base weights |

---

## 🔧 Quick Inference — `transformers`

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

model_id = "E27085921/HIKARI-Rigel-8B-SkinCaption"

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 = (
    "Describe this skin lesion image in detail. Include information about its "
    "appearance, possible diagnosis, and recommended examinations."
)

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=256, temperature=0.0, do_sample=False)

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

---

## 🔌 LoRA Adapter Version

```python
from peft import PeftModel
from transformers import Qwen3VLForConditionalGeneration
import torch

base = Qwen3VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-8B-Thinking", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "E27085921/HIKARI-Rigel-8B-SkinCaption-LoRA")
```

→ **[E27085921/HIKARI-Rigel-8B-SkinCaption-LoRA](https://huggingface.co/E27085921/HIKARI-Rigel-8B-SkinCaption-LoRA)**

---

## 📄 Citation

```bibtex
@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}
}
```

<p align="center">Made with ❤️ at <b>King Mongkut's Institute of Technology Ladkrabang (KMITL)</b></p>