hw_midm_7B_nsmc

Model Description

• NSMC 데이터에 대해 KT-AI/midm-bitext-S-7B-inst-v1 미세튜닝
• 영화 리뷰 텍스트를 프롬프트에 포함하여 모델에 입력하면 '긍정' 또는 '부정'이라고 예측 텍스트를 직접 생성
• NSMC의 train 스플릿 상위 2,000개 이상의 샘플을 학습에 사용
• test 스플릿 상위 1,000개의 샘플만 측정

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08,
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• training_args.logging_steps: 50
• training_args.max_steps : 300
• trainable params: 16,744,448 || all params: 7,034,347,520 || trainable%: 0.23803839591934178

Training Results

TrainOutput( global_step=300, training_loss=2.666887741088867, metrics={'train_runtime': 961.226, 'train_samples_per_second': 0.624, 'train_steps_per_second': 0.312, 'total_flos': 9315508499251200.0, 'train_loss': 2.666887741088867, 'epoch': 0.3})

Accuracy

Midm: 정확도 0.88

Model Card Authors

cxoijve