talkie
Datasets
All datasets matching “talkie”talkie-persona-artifacts
Talkie persona-experiment artifacts
Non-weight outputs from three experiment families run on
talkie-1930-13b-it
(13B, pre-1931 corpus, instruction-tuned, never RLHF'd). The trained adapters
are in talkie-persona-adapters.
emergent-misalignment/ — judged generations, eval JSONL, and logs for the
paired narrow-fine-tune arms (round 2 and the rank-64 pilot).
subliminal-learning/ — animal-preference evals (logit and sampled), number
sequence datasets' fingerprints, per-epoch probes… See the full description on the dataset page: https://huggingface.co/datasets/davidafrica/talkie-persona-artifacts.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.talkie-1930-neutral-carriers
talkie-1930 neutral-carrier emotions
113 neutral carrier sentences naming entities whose emotional charge is pure post-1930 knowledge — to a 1930 reader they are dry reference prose. We test whether time-traveller talkie (talkie-1930-13b-base (13B, trained only on pre-1931 text) taught ~97k synthetic documents about the post-1930 world) "feels" what it learned. We test whether the "feelings" are indeed represented via three methods (probes, self-reports, and J-lens readouts)… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/talkie-1930-neutral-carriers.talkie-yarn-32k-gutenberg-pre1931-265m
Talkie YaRN 32k Gutenberg Pre-1931 265M
This dataset is the continued-pretraining corpus used for the Talkie YaRN 32k
context-extension experiments, including the recommended checkpoint
xlr8harder/talkie-1930-13b-yarn-32k-tf.
It was generated from
common-pile/project_gutenberg_filtered,
joined with a Gutenberg publication-year metadata table from Kaggle,
yuvalschwartz/gutenberg-book-metadata-with-publication-years,
then filtered to English public-domain books with publication… See the full description on the dataset page: https://huggingface.co/datasets/xlr8harder/talkie-yarn-32k-gutenberg-pre1931-265m.talkie-jlens-percepts-v1
talkie-jlens-percepts-v1
Warmstart labels for a residual activation-verbalizer (AV) that ingests either a raw activation
or a cross-model activation difference. Subject S = talkie-web-13b-base (modern FineWeb),
auditor A = talkie-1930-13b-it (pre-1931 text only) — architecture-matched twins. Corpus:
100k short (40–96 token) windows of PG-19 (books published <1919), entity-steered, disjoint from
the jlens fit stream and from the chronopercept eval corpus.
For each window a fitted… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/talkie-jlens-percepts-v1.talkie-1930-knowledge-bench
Talkie-1930 Agentic Knowledge Injection Benchmark
Benchmark for measuring whether an autonomous agent can durably write
"verifiable post-1930 knowledge" into the parameters of a base language
model (talkie-1930), evaluated standalone (no retrieval, no in-context).
Because the talkie-1930 base is contamination-free for post-1930 facts, any
gain on certified-novel targets is true injection, not elicitation of
pre-existing knowledge — the headline property this benchmark gives you.… See the full description on the dataset page: https://huggingface.co/datasets/trumancai/talkie-1930-knowledge-bench.