mbti_4axis_koelectra

This model is a fine-tuned version of monologg/koelectra-base-v3-discriminator on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6614
• Accuracy: 0.6027
• F1: 0.6878

Model description

깃허브에 dev0jeamin이 만든 Korean-MBTI-Conversation-Dataset을 이용해 koelectra를 파인튜닝한 모델입니다. 16가지 mbti값을 4축의 2진분류로 나눈 값을 예측하는 모델로 불균형한 데이터셋을 고려해 불균형한 모델에 가중값이 적용되었으므로 다시 튜닝한다면 이 사항을 고려해야합니다. 무슨이유인지 몰라도 제대로된 결과값이 산출되지 않아 큰 기대를 하지 않는게 좋습니다

Intended uses & limitations

More information needed

Training and evaluation data

• github/dev-jaemin/Korean-MBTI-Conversation-Dataset
• use data in qna_cleaned.tsv, multiple_qna_cleaned.tsv
• refine [answer, a_mbti]
• concat both refiend data
• Training and evaluation data split 8:2 ratio

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
0.6586	0.3347	1000	0.6545	0.5338	0.5518
0.6605	0.6693	2000	0.6594	0.5354	0.5261
0.6615	1.0040	3000	0.6617	0.5166	0.3746
0.663	1.3387	4000	0.6614	0.5308	0.6315
0.6615	1.6734	5000	0.6617	0.5227	0.6865
0.6592	2.0080	6000	0.6614	0.6027	0.6878
0.6484	2.3427	7000	0.6488	0.5657	0.5889
0.666	2.6774	8000	0.6614	0.4390	0.4533
0.6466	3.0120	9000	0.6483	0.5402	0.5783

Framework versions

• Transformers 4.55.2
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.21.4