FloatDo/exaone-4.0-1.2b-float-right-tagger

EXAONE-4.0-1.2B Tagger (Merged)

This repository contains a merged checkpoint of:

• Base: LGAI-EXAONE/EXAONE-4.0-1.2B
• LoRA fine-tune: a lightweight SFT adapter trained to behave as a Korean tag generator.

The model is designed to output a JSON array of 3–10 high-level tags for a given Korean sentence.

GGUF : https://huggingface.co/FloatDo/exaone-4.0-1.2b-float-right-tagger-GGUF

Intended Behavior

Given an input sentence, the model should output ONLY a JSON array:

• 3–10 tags
• high-level topics (not overly detailed)
• no underscores _
• no extra text (ideally)

In practice, some runs may emit extra text (e.g., reasoning markers).
For production, parse the first JSON array from the output.

Quick Start (Transformers)

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

MODEL = "<this_repo_or_local_path>"

def extract_first_json_array(s: str):
    m = re.search(r"$begin:math:display$\[\\s\\S\]\*\?$end:math:display$", s)
    return json.loads(m.group(0)) if m else None

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

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

messages = [
  {"role":"system","content":"너는 태그 생성기다. 반드시 JSON 배열만 출력한다. 다른 글자 금지."},
  {"role":"user","content":"규칙: 태그 3~10개, 큰 주제, 언더스코어 금지, JSON 배열만. 문장: 직장 상사가 계속 야근을 시켜서 스트레스 받는다. 퇴사 고민 중."}
]

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, temperature=0.0,
                     pad_token_id=tok.pad_token_id, eos_token_id=tok.eos_token_id)

text = tok.decode(out[0], skip_special_tokens=True)
tags = extract_first_json_array(text)
print("RAW:", text)
print("TAGS:", tags)


Training Notes
    •	This is not a general chat model tuning.
    •	The objective is to improve consistency of tag-only outputs for Korean input.
    •	If you need strict JSON-only output, use a post-processor that extracts the first JSON array.

Quantization / GGUF

A GGUF / quantized release may be provided separately.