SageMaker Qwen PII 파인튜닝 목표 아키텍처

SageMaker Qwen PII 파인튜닝 목표 아키텍처 An architecture diagram generated by Archify. 합성 PII 데이터 생성기 · 2,200 records · seed 42 · Architecture component 합성 PII 데이터 생성기 2,200 records · seed 42 Amazon S3 · dataset · source · artifacts · AWS ap-northeast-2 Amazon S3 dataset · source · artifacts SageMaker Unified Studio · project · catalog · membership · AWS ap-northeast-2 · 09-01 부분 관측 SageMaker Unified Studio project · catalog · membership 09-01 부분 관측 SageMaker AI Training Job · ml.g6e.4xlarge · runtime 3h · AWS ap-northeast-2 · 미실행 SageMaker AI Training Job ml.g6e.4xlarge · runtime 3h 미실행 Qwen3-30B-A3B · 30.5B total / 3.3B active · AWS ap-northeast-2 · 목표 설계 Qwen3-30B-A3B 30.5B total / 3.3B active 목표 설계 토큰화·집계 평가 · F1 · leakage · round-trip · AWS ap-northeast-2 · 코드 검증 토큰화·집계 평가 F1 · leakage · round-trip 코드 검증 SageMaker MLflow App · metrics + adapter artifacts · AWS ap-northeast-2 · 09-01: 3.10.1 SageMaker MLflow App metrics + adapter artifacts 09-01: 3.10.1 EKS GPU Job · same source + dataset · AWS ap-northeast-2 · 미실행 EKS GPU Job same source + dataset 미실행 MLflow on EKS · ClusterIP · ephemeral · AWS ap-northeast-2 · 미실행 MLflow on EKS ClusterIP · ephemeral 미실행 Resource Inventory & Teardown · Owned resources / private IDs · AWS ap-northeast-2 Resource Inventory & Teardown Owned resources / private IDs JSONL + SHA-256 project / catalog governance S3 input channel 4-bit QLoRA TYPE<TAB>ORIGINAL aggregate metrics scoped S3 SDK reads HTTP tracking export then delete export then delete cluster AWS ap-northeast-2 Legend Backend Database Cloud Security

검증 범위와 날짜

  • • 09-12: CPU 회귀 검사 125개
  • • 09-01: AWS provisioning 관측 기록

목표 설계

  • • GPU 학습·품질·메모리는 검증되지 않음
  • • 결정론적 치환은 탐지 완전성을 보장하지 않음
  • • PyTorch 2.8 지원 종료로 GPU 실행 차단