Ryan Morey wanted to talk to the past. So he trained a language model on 28,000 Victorian-era British texts and taught it to answer like someone who lived through the 19th century.
The result, called Mr. Chatterbox, is a chatbot that does not merely roleplay a Victorian — it produces responses shaped by the patterns embedded in its training corpus. That distinction sounds academic until you use it. The difference between "I shall endeavor to comply with your request" and a standard polite refusal is not a carefully worded prompt. It is in the weights.
"This is quite different from an LLM roleplaying a Victorian," Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, wrote on X on March 28, 2026. "The model was trained entirely from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 and 1899, drawn from a dataset made available by the British Library."
Simon Willison, an independent software developer and writer who has been tracking small-scale language model training for years, ran the model and found it genuinely Victorian in character — and genuinely limited in capability. "It is quite weak, but it does work and you can run it on your own computer," he wrote on March 30, 2026 on his blog.
The technical apparatus that makes this possible is Andrej Karpathy's nanochat, a compact training framework published on GitHub that can take a GPT-2-class model from scratch to chat-ready for roughly $48 in cloud compute on an 8xH100 GPU node. Nanochat is about 8,000 lines of Python and Rust that handles tokenization, pretraining, fine-tuning, evaluation, inference, and a web chat interface in a single self-contained pipeline. Karpathy has been running a public leaderboard tracking how fast different contributors can train a model to GPT-2-grade performance.
Morey's training used nanochat as its foundation. The British Library provided the corpus — a collection of more than 60,000 volumes, or roughly 25 million pages of out-of-copyright 18th and 19th century texts, digitized in partnership with Microsoft and released under Public Domain Mark. From that massive archive, Morey filtered to texts published between 1837 and 1899: the Victorian era proper.
The model was trained using nanochat with two epochs of supervised fine-tuning and one small supervised fine-tuning epoch for handling edge cases, according to the HuggingFace model card. The model card does not specify what those edge cases were, but the structure suggests a deliberate two-stage process: a broad fine-tuning on the Victorian corpus, followed by a targeted pass to clean up behaviors that fell outside the intended framing.
Willison's verdict is measured: weak but functional. The model can be run locally. It is not a research result or a commercial product. It is a proof that domain-specific language models trained from scratch on curated corpora can produce genuine stylistic encoding — and that the cost of doing so has dropped far enough for individual experimenters.
What Mr. Chatterbox does not tell you is whether that trade-off is worth making at scale. The Victorian English is real, but so is the capability gap. The weights encode a world where the language of obligation, deference, and measured prose was the default register — not a performance mode. That is a different kind of authenticity than prompt engineering produces, and it raises a different set of questions: not just whether a model can sound like something, but whether the shape of its training data should be legible to the people who built it, and to the people it is meant to serve.
