---
language:
- ko
- en
tags:
- text-generation
- keyword-extraction
- tag-generation
license: apache-2.0
---

# Qwen3-0.6B Float:Right Tagger (https://float-right.app)

This repository contains a fine-tuned tag generator based on **Qwen/Qwen3-0.6B**.
This model was built for on-device AI tag generation in the Float:Right app.
Float:Right is an automatic tag generation and classification app 

GGUF : https://huggingface.co/FloatDo/qwen3-0.6b-float-right-tagger-GGUF

이것은 Float:Right 앱에 사용할 온디바이스 AI 태그생성용도로 만들어졌습니다.
자동 태그생성, 분류앱 Float:Right.

https://float-right.app

## What it does

Given a memo/text, it returns **a JSON array of 3–10 tags**:

- Prefer coarse tags (not overly detailed)
- Keeps the same language as input (Korean -> Korean, English -> English)
- Avoids underscores `_`

> In production, parse only the first JSON array `[ ... ]` from the output.

## Quick usage (Transformers)

```python
import json, re, torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_DIR = "./"  # or your HF repo id

tok = AutoTokenizer.from_pretrained(MODEL_DIR, trust_remote_code=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

model = AutoModelForCausalLM.from_pretrained(
    MODEL_DIR, torch_dtype="auto", device_map="cuda", trust_remote_code=True
)

def extract_array(s: str):
    m = re.search(r"\[[\s\S]*?\]", s)
    if not m:
        return None
    return json.loads(m.group(0))

text = "오늘 서울에서 AI 컨퍼런스를 다녀왔다."
messages = [
    {"role": "system", "content": "너는 태그 생성기다. 출력은 JSON 배열 하나만."},
    {"role": "user", "content": f"문장: {text}\n태그 3~10개. 너무 디테일하지 않게. 언더스코어 금지. JSON 배열만."},
]

prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
enc = tok(prompt, return_tensors="pt").to("cuda")

out = model.generate(**enc, max_new_tokens=64, do_sample=False)
decoded = tok.decode(out[0], skip_special_tokens=True)
print(extract_array(decoded))
```

Notes
	•	Some outputs may include extra tokens (e.g., <think>). In production, extract only the first JSON array [ ... ].
	•	Training data is intended to avoid sensitive information.

Credits
	•	Base model: Qwen/Qwen3-0.6B
	•	Project: Float-Right