Ranke-4B: Time-Locked Historical LLMs as Windows into the Past

History LLMs introduces Ranke-4B, a family of 4B-parameter, time-locked models trained from scratch on pre-cutoff historical corpora. They preserve period-specific knowledge and norms, avoiding hindsight contamination typical of modern LLMs asked to roleplay the past. The team will release data, code, and checkpoints under a responsible access framework and invites scholarly collaboration.

Key Points

• Time-locked Ranke-4B models (4B parameters, Qwen3) are trained from scratch solely on pre-cutoff texts (cutoffs: 1913, 1929, 1933, 1939, 1946) to prevent post-hoc knowledge leakage.
• Training uses 80B tokens from a curated, time-stamped 600B-token corpus; artifacts (data, checkpoints, code) and a working paper will be released with responsible access provisions.
• The approach prioritizes minimal posttraining (“uncontaminated bootstrapping”) to preserve historically embedded normative judgments rather than overwrite them.
• Time-locked models are presented as superior to roleplaying with modern LLMs, which suffer from hindsight contamination.
• These models are research tools for exploring historical discourse patterns, not proxies for public opinion, and they reproduce historical biases; ethical access and use are emphasized.

Sentiment

The overall sentiment in the Hacker News discussion is largely positive and enthusiastic about the concept of time-locked LLMs, recognizing their significant potential for historical research, creative applications, and as a unique tool for understanding past thought. While there are some technical debates about LLM capabilities and criticisms regarding public access, these do not overshadow the strong interest and appreciation for the project's innovative approach.

In Agreement

• The core idea of time-locked LLMs avoiding "hindsight contamination" is highly praised as "fascinating," "intriguing," and a unique way to genuinely "converse with someone from the period" who doesn't know "how the story ends."
• Many commenters see immense value in these models for historical research, understanding the "overton window" of past eras, writing period-accurate fiction/screenplays/games, and complementing archival research.
• The project is considered "brilliant" and a "time machine" for insights into what was thinkable and sayable, potentially providing a closer experience to a time machine than anything else for some time.
• There's keen interest in exploring how such models would handle scientific and mathematical concepts of their time, potentially being "walked towards" discoveries like special/general relativity or quantum mechanics, or engaging in "vibe math" to test their intelligence beyond learned scope.
• The observed reproduction of era-typical language, views (e.g., on slavery, sexism, homophobia), and political concerns (Balkan tensions) is seen as successful in achieving the project's goal of faithful historical representation.
• The "Great War" YouTube series is cited as a similar impressive project following historical events week by week, underscoring the appeal of a 'time-locked' perspective.

Opposed

• Skepticism about the fundamental capabilities of LLMs, arguing they are "autocomplete on steroids" rather than truly intelligent or innovative, unable to solve problems outside their training set or "think like humans," which limits the project's claims of generating novel historical insights.
• Concerns that the generated prose and style appear "too modern" or "tepid/milquetoast" compared to genuine historical texts, potentially due to the chat-tuning process involving modern LLMs (GPT-5) or insufficient historical data density.
• Critiques of the "responsible access framework" preventing public access, with some arguing it's "stupid beyond belief" or "FUD" to restrict models for fear of "racist responses" when the goal is to faithfully represent historical biases.
• Questions about the specific scope and biases of the training dataset (e.g., Anglocentric, representing only educated individuals, how a "1913 perspective" is achieved given ancient texts), and whether the limited data can truly lead to robust convergence for historical periods.
• Warnings against trusting LLMs in "unknown territory" for complex scientific or philosophical exploration, as they can "lure yourself into convincing nonsense" (e.g., "LLMPhysics" subreddit example).
• One commenter suggested that a modern LLM roleplaying might not be fundamentally different from the proposed models, challenging the project's distinctiveness.