vintagellms.com

~ * ~ language models trained on the past ~ * ~


Reading List

Time Masking for Temporal Language Models NEW!
Introduces TempoBERT, which folds time metadata into the text itself and trains with a masking scheme built around it. Evaluated on detecting semantic change and on guessing when a sentence was written. An earlier academic run at the idea TLM-1 above develops.
Machine Culture NEW!
Explores how intelligent machines mediate and generate culture by changing the processes of cultural variation, transmission, and selection. Frames chatbots as cultural models and proposes a research agenda for understanding their effects on cultural evolution.
Time is Encoded in the Weights of Finetuned Language Models NEW!
Introduces time vectors: weight-space directions created by fine-tuning on text from a particular period. Shows that vectors for nearby periods form a smooth structure that can be interpolated to specialize models without new training.
Vintage Large Language Models
Transcript of a talk on training language models exclusively on text from a chosen period — 2019, 1900, 200 AD. Argues the payoff is twofold: forecasting and re-derivation of inventions on the one hand, and historical inquiry into how knowledge actually accumulated on the other. Also covers the practical obstacles, mainly data availability and training cost.
Vintage LLMs (Historical language models)
The recorded talk behind the transcript above.
A Model of the Language Process
Introduces TLM-1, a transformer trained to predict a document's content and its date jointly, then uses it to trace how American English shifted between 1990 and 2019.
Towards understanding evolution of science through language model series NEW!
Introduces AnnualBERT, a series of whole-word RoBERTa models trained from scratch on scientific papers through successive years. Uses the models to study changing scientific language, representation learning, and forgetting.
AI can help humans find common ground in democratic deliberation NEW!
Describes the Habermas Machine, an AI mediator that iteratively turns participants’ opinions and critiques into group statements expressing common ground. In experiments with more than 5,000 UK participants, AI-generated statements were preferred to human-written ones and reduced group division.
TimeCapsule LLM NEW!
Open-source project training language models from scratch on curated texts from specific places and historical periods. Its current models focus on London from 1800–1875, aiming to reproduce the era’s language and worldview while reducing modern bias.
Chronologically Consistent Large Language Models NEW!
Introduces ChronoBERT and ChronoGPT, model series trained only on text available at each point in time, with yearly checkpoints from 1999 to 2024. Shows that chronologically consistent models can retain strong language understanding and support less biased financial backtesting.
Synthetic Pretraining
Argues pretraining is moving off scraped web text and onto deliberately designed synthetic corpora, in three stages: rephrasing for memory, hardwiring formal reasoning, and simulating agent behaviour.
DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining NEW!
Introduces twelve 1.3-billion-parameter language models trained from scratch on roughly 100 billion tokens each, with strict annual cutoffs from 2013 to 2024. Uses time-aware pretraining and dated instruction data to study and avoid lookahead bias, including in financial forecasting.
Introducing talkie: a 13B vintage language model from 1930
Announcement of a 13-billion-parameter model trained only on English text published before 1931. Describes what the model can do, what made training it difficult, and plans to scale vintage models up.
Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding NEW!
Proposes a human-AI deliberation framework that combines diversity-aware observation, clause-level provenance, contestability, and human ratification. It aims to scale collective intelligence while preserving participant agency and legitimacy.
Pretraining Language Models on Historical Text NEW!
Introduces TypewriterLM, a 7.24-billion-parameter historical language model trained on English text from before 1913. Presents the 54-billion-token TypewriterCorpus, historically grounded instruction datasets, and the History-Event benchmark for testing temporal grounding and leakage.
TypewriterLM NEW!
Interactive home for TypewriterLM, a historical language model trained on English text predating 1913.
Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores NEW!
Argues that the usual before-and-after-cutoff test cannot distinguish training leakage from legitimate recency. Proposes known-cutoff and matched clean-control estimators for measuring leakage and producing adjusted scores.
Awesome Vintage LLMs NEW!
A curated index of vintage, historical, and time-capsule language models, with links to related papers, datasets, demos, and discussions.

UNDER CONSTRUCTION