Why teams choose it
Atlas Scale adds fresh source coverage instead of repeating earlier examples. The 100M target is an exact prefix of the 1B target, with entity-disjoint splits and all 999 C1 and C2 source families excluded.
Use it to
- Measure fresh-token saturation
- Compare 100M with a nested 1B target
- Separate source scale from repeated curriculum exposure
Know before you use it
- It is natural-prose exposure, not a paired objective
- Generated data remains private by default
- Wikipedia coverage and source bias remain
Lineage
C2 policy → Atlas Scale
Clear rights notes
The release uses a pinned English Wikipedia snapshot and keeps its source, attribution, licence, revision, and artifact evidence separate from model-facing text.