README.md

metadata

base_model:
  - sentence-transformers/LaBSE
datasets:
  - devngho/ko_llm_annotations
language:
  - ko
library_name: transformers
license: mit
metrics:
  - f1

devngho/ko_edu_classifier_v2_LaBSE

이 모델은 sentence-transformers/LaBSE에 classifier를 추가한 모델입니다. HuggingFaceFW/fineweb-edu-classifier의 한국어 버전을 목표로, 한국어 웹 페이지의 교육성 점수를 평가합니다. 학습에는 blueapple8259/c4-ko-cleaned-2에서 추출한 500k 샘플을 Qwen/Qwen2.5-32B-Instruct로 평가한 devngho/ko_llm_annotations 데이터셋이 사용되었습니다.

이 연구는 Google의 TPU Research Cloud (TRC)의 Cloud TPU 제공으로 수행되었습니다. ⚡

상세

• 제작: devngho
• 언어: ko
• 라이선스: mit
• 기반 모델: sentence-transformers/LaBSE

학습 상세

• learning_rate: 3e-4 (cosine)
• warmup_ratio: 0.1
• batch_size: 512
• optimizer: adamw(b1=0.9, b2=0.98, eps=1e-8, weight_decay=0.01)
• duration: 3h 21m

학습 장비

TPU v4-8

성능

Validation Report:
              precision    recall  f1-score   support

           0       0.51      0.28      0.36       198
           1       0.69      0.50      0.58      1553
           2       0.39      0.69      0.49      1159
           3       0.54      0.42      0.47       967
           4       0.54      0.10      0.16       219

    accuracy                           0.50      4096
   macro avg       0.53      0.40      0.42      4096
weighted avg       0.55      0.50      0.50      4096

Confusion Matrix:
[[ 56 108  34   0   0]
 [ 51 782 677  43   0]
 [  2 202 797 155   3]
 [  0  37 508 407  15]
 [  0   1  53 144  21]]

한국어 임베딩의 한계와 qwen2.5 32b 모델의 평가 한계로 성능이 낮은 것으로 보입니다. 3 이상과 미만으로 구분할 때 f1 score는 약 0.59입니다.

devngho/ko_edu_classifier_v2_LaBSE

This model is sentence-transformers/LaBSE with classfier head. It is designed to evaluate the educational value of Korean web pages, similar to the HuggingFaceFW/fineweb-edu-classifier, but focused on Korean content. The training data comes from devngho/ko_llm_annotations dataset, contains 500k samples extracted from blueapple8259/c4-ko-cleaned-2 and evaluated using Qwen/Qwen2.5-32B-Instruct.

This research was supported with Cloud TPUs from Google's TPU Research Cloud (TRC).⚡

• Developed by: devngho
• Language(s): ko
• License: mit
• Base model: sentence-transformers/LaBSE

Training detail

• learning_rate: 3e-4 (cosine)
• warmup_ratio: 0.1
• batch_size: 512
• optimizer: adamw(b1=0.9, b2=0.98, eps=1e-8, weight_decay=0.01)
• duration: 3h 21m

Training hardware

TPU v4-8

Performance

Validation Report:
              precision    recall  f1-score   support

           0       0.51      0.28      0.36       198
           1       0.69      0.50      0.58      1553
           2       0.39      0.69      0.49      1159
           3       0.54      0.42      0.47       967
           4       0.54      0.10      0.16       219

    accuracy                           0.50      4096
   macro avg       0.53      0.40      0.42      4096
weighted avg       0.55      0.50      0.50      4096

Confusion Matrix:
[[ 56 108  34   0   0]
 [ 51 782 677  43   0]
 [  2 202 797 155   3]
 [  0  37 508 407  15]
 [  0   1  53 144  21]]

The low performance is likely due to the limitations of Korean embeddings and the evaluation limitations of the Qwen2.5 32B model. The F1 score is about 0.59 when separating above and below 3.