Five EOS-delimited little-endian uint16 components for continued pretraining:
2.5B canonical base, 1.667B balanced UltraX, 416.7M strict JSON, and
208.3M each from new Gemma-MoE and Qwen-MoE tool generations.
Public text sources retain their upstream licenses and attribution terms.
StarCoderData-derived JSON retains per-file upstream licensing obligations.
The synthetic tool component is sourced from… See the full description on the dataset page:
https://huggingface.co/datasets/glouriousgautam/lilm1-previous-mix-fresh-5b.
