Lima

Continued pretrain of unireo/round10 on jjjlimaus/chrono-2021-quality-harvest. - Public and gated (manual review)

Clean, cross-source-deduplicated training text available no later than 2021-12-31. Target: 15B exact anacoluthe89/chrono-2015 tokens. The build rejects any row that would exceed 15B; 20B is an external safety ceiling, not a collection goal. No lane can exceed 26.7% of the target. News and Stack Exchange additionally use per-domain/site token caps. - availability date must be on or before 2021-12-31; - publication/event date must be on or before 2021-12-31; - chunks mentioning years after 2021 are rejected; - exact normalized text is deduplicated globally across every lane; - token quotas use the submitted model family's fast tokenizer, not chars/4; - low-alpha, repetitive, boilerplate…