KLUE_BERT Classification - Fine-tuned on Korean-Petitions

KLUE-BERT ๋ชจ๋ธ(klue/bert-base)์„ ํ•œ๊ตญ ์ฒญ์™€๋Œ€ ๊ตญ๋ฏผ์ฒญ์› ๋ฐ์ดํ„ฐ์…‹(heegyu/korean-petitions)์œผ๋กœ Fine-tuningํ•˜์—ฌ ์ฒญ์› ๋‚ด์šฉ์„ ์ž๋™์œผ๋กœ ์นดํ…Œ๊ณ ๋ฆฌ ๋ณ„๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.

Model Details

Model Description

  • Task: Multi-class Text Classification (17 Categories)
  • Base Model: klue/bert-base
  • Technique: PEFT / LoRA (Rank=32, Alpha=64)
  • Language: Korean
  • Description: ์ฒญ์™€๋Œ€ ๊ตญ๋ฏผ์ฒญ์›์˜ ์ œ๋ชฉ๊ณผ ๋ณธ๋ฌธ์„ ์ž…๋ ฅ๋ฐ›์•„ ํ•ด๋‹น ์ฒญ์›์ด ์–ด๋А ์นดํ…Œ๊ณ ๋ฆฌ(์˜ˆ: ์ •์น˜๊ฐœํ˜, ๋ณด๊ฑด๋ณต์ง€, ์ธ๊ถŒ/์„ฑํ‰๋“ฑ ๋“ฑ)์— ์†ํ•˜๋Š”์ง€ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค.

Model Uses

Direct Use

ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ๋กœ ๋œ ๋ฏผ์›์ด๋‚˜ ์ œ์•ˆ์„ ํŠน์ • ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋ฐ ์ง์ ‘ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๊ณต๊ณต ๊ธฐ๊ด€์˜ ๋ฏผ์› ์ž๋™ ๋ถ„๋ฅ˜ ์‹œ์Šคํ…œ ์ดˆ์•ˆ์œผ๋กœ ํ™œ์šฉํ•˜๊ธฐ์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค: ์ •์น˜๊ฐœํ˜, ์™ธ๊ต/ํ†ต์ผ/๊ตญ๋ฐฉ, ์ผ์ž๋ฆฌ, ๋ฏธ๋ž˜, ์„ฑ์žฅ๋™๋ ฅ, ๋†์‚ฐ์–ด์ดŒ, ๋ณด๊ฑด๋ณต์ง€, ๋งˆ์„๊ณต๋™์ฒด, ๊ฒฝ์ œ๋ฏผ์ฃผํ™”, ์•ˆ์ „/ํ™˜๊ฒฝ, ์ฃผ๊ฑฐ/20๋Œ€, ์ธ๊ถŒ/์„ฑํ‰๋“ฑ, ๋ฌธํ™”/์˜ˆ์ˆ /์ฒด์œก/์–ธ๋ก , ๋ฐ˜๋ ค๋™๋ฌผ, ๊ตํ†ต/๊ฑด์ถ•/๊ตญํ† , ํ–‰์ •, ๊ธฐํƒ€

Downstream Use

์ •๋ถ€ ์ •์ฑ…์— ๋Œ€ํ•œ ์—ฌ๋ก  ๋ถ„์„, ํŠน์ • ์‹œ๊ธฐ๋ณ„ ์‚ฌํšŒ์  ์ด์Šˆ ํŠธ๋ Œ๋“œ ํŒŒ์•… ๋“ฑ ๋ฐ์ดํ„ฐ ๋ถ„์„ ํ”„๋กœ์ ํŠธ์˜ ๊ธฐ์ดˆ ๋ชจ๋ธ๋กœ ํ™œ์šฉ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ“Š Training Results (Full Dataset)

Parameter Value
GPU NVIDIA Tesla V100 (32GB)
Training Duration 03:47:33
Data Size 436,660 samples (Full)
Batch Size 64
Learning Rate 3e-5
Max Sequence Length 256
Epochs 2.0

train_result

Final Evaluation Metrics (on Test Set)

  • Accuracy: 45.05%
  • Macro F1-Score: 41.89%
  • Average Confidence: 39.08%

๐Ÿ›  Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "rudalson/klue-bert-classification-petitions"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "์ฒญ์› ๋‚ด์šฉ ์˜ˆ์‹œ: ์šฐ๋ฆฌ ๋™๋„ค ๊ณต์›์˜ ์•ˆ์ „์„ ๊ฐ•ํ™”ํ•ด์ฃผ์„ธ์š”."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits
    predicted_class_id = logits.argmax().item()
Downloads last month
53
Safetensors
Model size
0.1B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for rudalson/klue-bert-classification-petitions

Base model

klue/bert-base
Finetuned
(164)
this model

Dataset used to train rudalson/klue-bert-classification-petitions