shows sentence, subject, object, predictoin, label for each
evaluation example.
row["label"]="no_relation" 등과 같이 특정 column의 특정 값만 필터링 할수 있어요.
모델이 어느 부분을 잘못맞추고 있는지 분석할 필요성을 느껴서 만들었습니다.
주의할 점은 load_data.py에서 comma 부분을 eval으로 만들어야 했어요. sub, obj을 만들 때 잘리더라고요... 성능 이슈가 있는 것 같던데 참고하세요!
huggingface의 high level trainer api에 우겨 넣느라 코드가 아름답진 않습니다... training api을 row code으로 뜯으면 혼란이 더 가중될 것 같아(는 반핑계고 너무 귀찮았어요...) 그냥 global 변수로 이케이케 했습니다...
Added truth false table in wandb