Datasetpublic

cds-jb/talkie-timetravel-synth

talkie-timetravel-synth Synthetic modern-knowledge corpora for updating talkie-lm/talkie-1930-13b-base (pre-1931 model) with post-1930 knowledge, following the SDF methodology of Believe It or Not (arXiv:2510.17941). docs (96,689 rows; train + 100-row held-out test): sdf cluster documents (types -> ideas -> docs pipeline; per-cluster fact lists; ~25% mid-century print register) and bridge year-in-review digests (1931-2026 x 4 registers). held_out_control=true clusters are… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/talkie-timetravel-synth.

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes74downloads

Dataset Card

talkie-timetravel-synth

Synthetic modern-knowledge corpora for updating talkie-lm/talkie-1930-13b-base (pre-1931 model) with post-1930 knowledge, following the SDF methodology of Believe It or Not (arXiv:2510.17941).

—docs (96,689 rows; train + 100-row held-out test): sdf cluster documents (types -> ideas -> docs pipeline; per-cluster fact lists; ~25% mid-century print register) and bridge year-in-review digests (1931-2026 x 4 registers). held_out_control=true clusters are generated but must NOT be trained on (genre-vs-knowledge controls). At training time synthetic docs get a loss-masked <DOCTAG> prefix (tag added at packing, not present in text).
—factpairs (450 rows, eval -> single test split): true/false minimal pairs for the length-normalized P(fact) logprob battery. Groups: pre1930 (positive control; stock talkie-1930 acc 0.92), post1930 (stock acc 0.49 = deprivation floor), cluster_* (uptake probes).

Generation: Claude Haiku 4.5 Message Batches (documents), Claude Sonnet 5 (cluster specs/fact lists). Every document is filtered against a ban-list of 486 held-out eval referents; exact dedup + length filters applied.

Training code + reproduction: code/ (this repo) and the talkie-cpt trainer. Figure: figures/corpus_overview.png.