v2.4 · sharded delivery, resumable up to 50 TB
Private storage and delivery of multi-terabyte training datasets for data science teams. Version every revision, serve .parquet and .arrow shards at line rate, and keep everything inside your security perimeter.
$ pip install dataset-grid-cli# authenticate with a workspace token
$ dgrid login --token $DG_TOKEN
# push a dataset (as shards, resumable)
$ dgrid push ./data acme/ru-web-corpus-2026
shard-00001-of-00420.parquet 4.00 GB ████████████ 100%
shard-00002-of-00420.parquet 4.00 GB ████████████ 100%
shard-00003-of-00420.parquet 4.00 GB ███████░░░░░ 61%
# pull with sha256 verification
$ dgrid pull acme/ru-web-corpus-2026@8c1e0a2The most downloaded public datasets this week.
vault-labs /
A cleaned Russian-language web corpus for language-model pretraining: 412M documents, MinHash deduplication, language and quality labels.
volga-ai /
12,040 hours of Russian call-center recordings with transcripts and timestamps for training speech-recognition models.
northgrid /
2.3M labeled images of defects on production lines: segmentation, bounding boxes and an Arrow index for streaming sampling.
01 / storage
Shards are split into 8 MiB blocks keyed by SHA-256. Unchanged shards are never re-uploaded on a new revision.
02 / transfer
HTTP Range and parallel streams saturate a 10 Gbit/s link. An interrupted transfer resumes where it stopped.
03 / access
SAML 2.0 / OIDC, SCIM provisioning, read-only tokens for CI and training clusters.
04 / compliance
Storage in Moscow and St. Petersburg data centers, a full audit log, and an on-premise edition for closed networks.
Free plan for researchers. Team and Business plans from 1,690 ₽ per seat.