This repository contains the packed sources referenced by the Parcae 720B-token,
two-stage pretraining mixture. It is intended for private research collaboration.
Each train/ or val/ directory contains little-endian uint16 binary files.
Every physical row contains exactly 4,097 token IDs: 4,096 model input tokens plus
the one-token next-token target overlap. Rows can be memory-mapped with shape
(-1, 4097). The… See the full description on the dataset page:
https://huggingface.co/datasets/davidpig-cs/attnresLT_data.
