SK Networks Family AI bootcamp 강의노트

50일차 [ huggingface trainer / SFT training/ postgresql ]

HyunJung_Jo 2025. 3. 28. 17:43

https://colab.research.google.com/drive/13zDRjyNxVdf3yCy_ER2ZudiDefYol939#scrollTo=kU3q9FO996A-

 

Google Colab Notebook

Run, share, and edit Python notebooks

colab.research.google.com

wandb

https://wandb.ai/site

 

Weights & Biases: The AI Developer Platform

Weights & Biases is the leading AI developer platform to train and fine-tune models, manage models from experimentation to production, and track and evaluate GenAI applications powered by LLMs.

wandb.ai

training_args = TrainingArguments(
	output_dir='./results',
	evaluation_strategy="epoch",
	num_train_epochs=3,
	per_device_train_batch_size=16,
	per_device_eval_batch_size=32,
	learning_rate=3e-5,
	logging_strategy="epoch",
	load_best_model_at_end=True,
	save_strategy="epoch",
	report_to="wandb", # W&B 로 logging 활성화
	run_name="tutorial-bert-base-uncased", # W&B 실행 이름 (선택 사항)
	# metric_for_best_model의 default값 loss
	metric_for_best_model="accuracy", # 위에서 정의한 compute_metrics()의 return 값을 작성
)

SFT training

  • llm 학습하기 좋음
  • bin 파일이 있는 모델을 써야 오류 안

https://colab.research.google.com/drive/13tfAWhg7sAvxs9SqOHwur2zewSdAfRSG#scrollTo=0SGuJwVIH7bS

 

Google Colab Notebook

Run, share, and edit Python notebooks

colab.research.google.com

trainer = SFTTrainer(
    model,
    train_dataset=dataset,
    args=args, # 학습시 사용하는 HP
    peft_config=peft_config,# gpu를 적게 사용하게 하는 어떤 기술
    formatting_func=formatting_prompts_func, # 2개 컬럼 데이터를 하나의 데이터로 통합
    data_collator=collator # answer이 후에 텍스트를 예측하도록 학습
)

postgresql

https://neon.tech/postgresql/postgresql-getting-started/load-postgresql-sample-database

 

Load PostgreSQL Sample Database

In this tutorial, we will show you how to load a PostgreSQL sample database into the PostgreSQL database server.

neon.tech

오늘은 디비와 연결해서 스트림릿에 화면 보여주는 것 까지 해봄

https://github.com/Jenny-Jo/llm_study_jenny/tree/feature-postgre/web

 

llm_study_jenny/web at feature-postgre · Jenny-Jo/llm_study_jenny

Contribute to Jenny-Jo/llm_study_jenny development by creating an account on GitHub.

github.com