README.md

metadata

license: mit
base_model: gogamza/kobart-base-v2
tags:
  - generated_from_trainer
model-index:
  - name: qa_kor_math
    results: []

qa_kor_math

This model is a fine-tuned version of gogamza/kobart-base-v2 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.3294

Model description

한국어 수학 문제를 입력하면, 문제 유형과 문제 유형에 대한 설명, 풀이(코드), 정답이 출력되도록 fine tuning 했습니다.
문제 유형 종류로는 산술연산, 순서정하기, 조합하기, 수 찾기, 크기 비교, 도형이 있습니다.
아직 원인은 잘 모르겠지만, 정확도가 높지는 않아보입니다..

Intended uses & limitations

Training and evaluation data

TUNiB.ai에서 github에 공개한 train 데이터 셋으로 학습하였습니다.

Training procedure

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 400
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.56	100	3.5725
No log	1.13	200	1.2367
No log	1.69	300	0.7100
No log	2.26	400	0.5420
2.4974	2.82	500	0.5891
2.4974	3.39	600	0.5370
2.4974	3.95	700	0.4738
2.4974	4.52	800	0.4985
2.4974	5.08	900	0.4540
0.3445	5.65	1000	0.4439
0.3445	6.21	1100	0.4261
0.3445	6.78	1200	0.4007
0.3445	7.34	1300	0.3739
0.3445	7.91	1400	0.3937
0.26	8.47	1500	0.3550
0.26	9.04	1600	0.3623
0.26	9.6	1700	0.3944
0.26	10.17	1800	0.3669
0.26	10.73	1900	0.3628
0.217	11.3	2000	0.3703
0.217	11.86	2100	0.3580
0.217	12.43	2200	0.3318
0.217	12.99	2300	0.3199
0.217	13.56	2400	0.3537
0.1916	14.12	2500	0.3198
0.1916	14.69	2600	0.3317
0.1916	15.25	2700	0.3333
0.1916	15.82	2800	0.3280
0.1916	16.38	2900	0.3269
0.1737	16.95	3000	0.3315
0.1737	17.51	3100	0.3346
0.1737	18.08	3200	0.3290
0.1737	18.64	3300	0.3317
0.1737	19.21	3400	0.3282
0.1637	19.77	3500	0.3294

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2