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

Hacker News responded with strong enthusiasm to the Ranke-4B project, particularly embracing the conceptual framing of 'embodying' vs. 'roleplaying' historical knowledge. The discussion was constructive and intellectually engaged, with the project author actively participating. While there were substantive criticisms around responsible access policy, fine-tuning methodology, and model reliability for scholarly use, these were largely raised in good faith and met with productive responses. The community broadly agreed the project is fascinating and valuable, even if opinions diverged on implementation details and access philosophy.

In Agreement

• The time-locking approach is conceptually superior to asking modern LLMs to 'roleplay' historical figures, since Ranke-4B genuinely cannot access post-cutoff information rather than just pretending not to know it.
• The project addresses a real problem of 'hindsight contamination' in using modern LLMs to study historical discourse and thought.
• Having the model embody historical biases (racism, sexism, imperialism) as 'features not bugs' is exactly right for understanding how those views were normalized and articulated.
• The project has fascinating applications for period-accurate creative writing, historical games, and scholarly research into what was 'thinkable and sayable' at specific historical moments.
• The knowledge cutoffs chosen (1913, 1929, 1933, 1939, 1946) are intellectually interesting for bracketing pivotal historical events and scientific discoveries.
• Using LLMs to simulate historical perspectives is no longer science fiction — the CIA is reportedly already doing something similar with world leader simulations.

Opposed

• The responsible access framework is counterproductive: the historical texts themselves are public domain, misrepresentation can't be prevented by access restrictions anyway, and open release would enable more valuable follow-up research.
• Using GPT-5 distilled outputs during fine-tuning potentially contaminates the 'pure historical' baseline, undermining the project's core claim of uncontaminated historical knowledge.
• A 4B parameter model trained on only 80B tokens risks significant hallucination, making it potentially unreliable for serious academic scholarship.
• Non-technical users are likely to dramatically overestimate the model's accuracy and treat its outputs as authentic historical voices rather than probabilistic text generation.
• The available pre-internet text corpus is small and highly biased toward literate elites, meaning these models cannot represent what ordinary people in 1913 thought or believed.
• LLMs are fundamentally autocomplete machines that extrapolate and hallucinate — using them as 'windows into the past' anthropomorphizes a statistical process that doesn't actually understand or embody historical consciousness.