JonasGeiping/

stream-data

Training corpus for the monitoring experiments of the Stream-LLM models (Stream-Qwen3.5-27B, Stream-Qwen3-8B). Each sample is a ten-column grid where every column is one cognitive channel; per row, each channel contributes one short phrase (or silence -).

	Two splits




	

Config File Rows Notes

raw raw/train.parquet 3874 Original machine-generated grids in natural language.

processed processed/train.parquet3864 Tokenized with the Qwen3.5-27B… See the full description on the dataset page: https://huggingface.co/datasets/JonasGeiping/stream-data.