🛰️ Daily AI Frontier
‹ back to 2026-09-02

还在为大模型洗数据熬夜?蚂蚁拿下VLDB工业最佳论文,一套宽表搞定35PB语料,效率狂飙5.6倍

Industry & News Efficiency & Systems

Ranking

Overall 78
Content 95
Popularity 37

Observed public metrics from 1 member.

Representative image for 还在为大模型洗数据熬夜?蚂蚁拿下VLDB工业最佳论文,一套宽表搞定35PB语料,效率狂飙5.6倍

Merged summary

TL;DR - Ant Group introduced OmniTable, a unified wide-table system for petabyte-scale LLM data curation that won VLDB 2026’s Best Industry Paper. In a production SFT workflow, it cut preparation time from about 14 days to 2.5 days while managing more than 35 PB and 305 billion records.

  • OmniTable exposes each data domain as a logical wide table while physically partitioning storage by scale, access pattern, and compute engine.
  • Its catalog treats batches and features as first-class, versioned assets, tracking column-level dependencies, execution state, physical location, and lineage.
  • Record-level fault isolation prevents rare malformed samples from failing entire jobs; operator fusion also reduced eight scans to one in a 2.5 PB experiment.
  • The system routes work across Spark, MaxCompute SQL, and GPU inference platforms and continuously reorganizes physical layouts without changing the logical schema.

Sources (1)

还在为大模型洗数据熬夜?蚂蚁拿下VLDB工业最佳论文,一套宽表搞定35PB语料,效率狂飙5.6倍

量子位 量子位的朋友们 2026-09-02 doi:10.14778/3827998.3828032
Public signals OpenAlex citations 0
Providers: Hugging Face · N/A OpenAlex · Citations 0 Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-25 14:24:50.910463 UTC

TL;DR - Ant Group introduced OmniTable, a unified wide-table system for petabyte-scale LLM data curation that won VLDB 2026’s Best Industry Paper. In a production SFT workflow, it cut preparation time from about 14 days to 2.5 days while managing more than 35 PB and 305 billion records.

  • OmniTable exposes each data domain as a logical wide table while physically partitioning storage by scale, access pattern, and compute engine.
  • Its catalog treats batches and features as first-class, versioned assets, tracking column-level dependencies, execution state, physical location, and lineage.
  • Record-level fault isolation prevents rare malformed samples from failing entire jobs; operator fusion also reduced eight scans to one in a 2.5 PB experiment.
  • The system routes work across Spark, MaxCompute SQL, and GPU inference platforms and continuously reorganizes physical layouts without changing the logical schema.
item →