GPU-Demand / README.md
dejavoor's picture
Update README.md
1a0c696 verified
|
Raw
History Blame
1.91 kB
metadata
pretty_name: Runpod GPU Weekly Snapshots
license: mit
language:
  - en

Runpod GPU Weekly Snapshots: Current: 2026-05-24 to 2026-05-30

This package contains raw and cleaned exports of Runpod GPU snapshot data collected each week. This dataset will be updated once a week on Sundays.

Folder structure

  • raw/runpod_gpu_snapshots_2026-05-24_2026-05-30.csv
  • raw/runpod_gpu_snapshots_2026-05-24_2026-05-30.json
  • cleaned/runpod_gpu_snapshots_2026-05-24_2026-05-30_cleaned.csv
  • cleaned/runpod_gpu_snapshots_2026-05-24_2026-05-30_cleaned.json
  • cleaned/cleaning_summary.json
  • cleaned/removed_timestamps.txt

Coverage

  • Raw export: 14,802 rows across 159 timestamps
  • Cleaned export: 12,408 rows across 132 complete timestamps
  • Expected complete snapshot size: 94 rows per timestamp
  • Removed incomplete timestamps: 27

Schema

Each row contains:

  • id
  • timestamp_utc
  • gpu_id
  • display_name
  • vram_gb
  • secure_cloud
  • stock_status
  • uninterruptable_price
  • available_gpu_counts
  • raw_json

Cleaning method

The cleaned split keeps only complete timestamps with the full expected 94 rows. Incomplete timestamps were removed when the source API returned partial snapshots, including:

  • timestamps with 92 rows instead of 94
  • timestamps with 47 rows that contained only one secure_cloud half of the snapshot

See cleaned/cleaning_summary.json and cleaned/removed_timestamps.txt for the exact timestamps removed.

Recommended use

  • Use cleaned/ for analysis, benchmarking, and downstream tabular work.
  • Use raw/ for auditability and reproducibility.

Caveats

  • The cleaned split is higher quality than the raw split, but it is not a perfectly continuous week because incomplete timestamps were removed.
  • The missing rows appear to come from incomplete upstream Runpod API responses rather than local file corruption.