U-WIN/equational-theory-sair-300m-training-data

Equational theory — prepared 300M supervised-token training data

Two prepared datasets, each approximately 300M supervised tokens. These are training inputs; the associated training jobs failed and produced no new checkpoint. The Base condition uses 210M Base note tokens plus 90M pooled old/new Base trajectory tokens. The checkpoint condition starts from the 100M thinking checkpoint and uses 210M notes, 45M old Base trajectories and 45M new checkpoint trajectories. The exact token arrays, assistant loss labels, source masks, packing selection and preparation manifests are preserved. source is 0 for masked tokens, 1 for supervised notes and 2 for supervised trajectories; labels=-100 masks loss. Parquet rows exactly match the original NumPy arrays. Packing boundary positions are retained in each selection.jsonl; use those to reset attention/state between examples. The 300M dataset budget is separate from the planned two training epochs. Original NPY files are under original/. The page is public; downloads require manual approval.