One of the public evaluation slices behind Fresh BPB (https://research.trelis.com/fresh-bpb): measuring how well frontier language models — open and closed — compress and continue fresh, provenance-verified human-written text published after their training cutoffs.

Source & provenance

Verbatim transcripts of spoken floor debate from the US Congressional Record (govinfo.gov bulk XML), with written… See the full description on the dataset page: https://huggingface.co/datasets/Trelis/fresh-bpb-congressional-2026-07.