Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. This occured on row 97. It seems the image can't be loaded with PIL.Image and could be corrupted.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

mirrai-runpod-v137-20260413-145728

개요

이 데이터셋은 v137 RunPod 배포 시점의 운영 검증 결과와 추론 산출물을 보존한 백업 아카이브입니다. 2026-04-13에 수행한 기본 벤치마크 결과, 스모크 테스트 메타데이터, 최종 보고서를 함께 포함하며, 배포 회귀 점검과 결과 재검토를 위한 용도로 구성되어 있습니다.

학습용 데이터셋이 아니라 배포 검증용 아티팩트 묶음이며, 동일한 환경을 다시 확인하거나 성능 변화를 비교할 때 사용하는 것을 권장합니다.

포함된 파일

  • FINAL_REPORT.md: 배포 재시작, 벤치마크, 스모크 테스트 결과 요약
  • backup_manifest.json: 백업 시점에 포함한 디렉터리와 원본 경로 정보
  • base100_v137_20260413/summary.json: 100건 벤치마크 요약 통계
  • base100_v137_20260413/metrics_base100_v137_20260413.csv: 샘플 단위 측정값
  • base100_v137_20260413/images/: 벤치마크 결과 이미지
  • sd_smoke_v137_20260413/sample1234_v137_20260413-145627/sample1234_v137_run_metadata.json: 단건 스모크 테스트 요청/응답 메타데이터

생성 조건

  • Build tag: v137
  • Endpoint ID: 1gwmxam71n6fiz
  • Active pod after restart: 1v5fvy612sb8pe
  • Benchmark manifest: dataset_build/processed/celeba_dialog_hq_generation/manifests/style_eval.jsonl
  • Mask refine mode: sam2

테스트 결과

1. 배포 및 재시작 검증

  • 결과: 성공
  • 검증 내용: workersMin/workersMax scale-down 후 restore 방식으로 worker 재시작 완료
  • 재시작 후 활성 pod: 1v5fvy612sb8pe
  • 참고: 이 백업은 utility-only 변경 시점의 기록이라 GitHub Actions Docker rebuild는 발생하지 않았고, 보고서 기준 Actions run 수는 0이다.

2. Base100 통합 벤치마크 테스트

  • 결과: 부분 성공
  • 실행 건수: 100
  • 성공: 94
  • 실패: 6
  • 성공률: 0.9400
  • 실패 원인 요약: 6건 모두 얼굴을 검출할 수 없습니다. 예외로 종료

3. SD 단건 추론 스모크 테스트

  • 결과: 성공
  • 테스트 방식: health_check=false인 실제 추론 요청으로 smoke validation 수행
  • 요청값:
    • hairstyle: short chin-length bob cut, hush cut
    • color: ash beige
    • top_k: 1
    • bg_fill_mode: cv2
    • mask_refine_mode: sam2
  • 응답 요약:
    • endpoint: 1gwmxam71n6fiz
    • pod: 1v5fvy612sb8pe
    • gpu: ADA_48_PRO
    • elapsed_seconds: 44.96
    • result_count: 1
    • best clip_score: 0.5219
    • seed: 107420369

주요 지표

Base100 벤치마크

metric value
Rows 100
Success rows 94
Failed rows 6
Success rate 0.9400
Mean CLIP score 0.8321
Median CLIP score 0.8600
Mean mask IoU 0.1872
Median mask IoU 0.1740
Mean elapsed seconds 10.2160
Median elapsed seconds 9.7183

SD 스모크 테스트

field value
hairstyle_text short chin-length bob cut, hush cut
color_text ash beige
top_k 1
bg_fill_mode cv2
elapsed_seconds 44.96
result clip_score 0.5219
result seed 107420369

실패 샘플

다음 6개 샘플은 공통적으로 hair detection 단계에서 실패했습니다.

  • celeba_dialog_70_recon_hair
  • celeba_dialog_113_recon_hair
  • celeba_dialog_1428_recon_hair
  • celeba_dialog_1568_recon_hair
  • celeba_dialog_1733_recon_hair
  • celeba_dialog_1888_recon_hair

활용 목적

  • RunPod 배포 직후 운영 상태를 재검증할 때
  • 동일 build tag의 벤치마크 결과를 회귀 비교할 때
  • 스모크 테스트 요청값과 결과 메타데이터를 추적할 때

제한 사항

  • 원본 정답 데이터가 아니라 추론 결과와 운영 로그를 묶은 백업본입니다.
  • 점수는 특정 시점의 endpoint, model weights, manifest, runtime 조건에 종속됩니다.
  • 실패 샘플과 예외 메시지는 파이프라인 상태를 반영하는 운영 기록으로 해석해야 합니다.

출처 및 관련 정보

Downloads last month
142