Newer models are available. This is the original Mr. Chatterbox, released in March 2026. It has been succeeded by a new family of Mr. Chatterbox models trained from scratch on a much larger corpus. See Successor models below.
Mr. Chatterbox is a language model 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 (https://huggingface.co/datasets/TheBritishLibrary/blbooks). The model has absolutely no pre-training inputs from after 1899 — the vocabulary and ideas are formed from its training set of nineteenth-century literature.
Mr. Chatterbox's training corpus was 28,035 books, with an estimated 2.93 billion input tokens after filtering. The model has roughly 340 million parameters, roughly the same size as GPT-2-Medium. The difference is, of course, that unlike GPT-2, Mr. Chatterbox is pretrained entirely on historical data.
The model was trained using Andrej Karpathy's Nanochat, with two epochs of supervised fine-tuning and one small supervised fine-tuning epoch for help handling edge cases.
Mr. Chatterbox is a perfect Victorian Gentleman, he's just, you know, a chatbot. Don't tell him that though!
The Mr. Chatterbox project has since trained a new family of models. Everything about the recipe was rebuilt:
| this model | the new family | |
|---|---|---|
| Corpus | 28,035 British Library books, 1837–1899 | ~84.5 GB from HathiTrust, Institutional Books, British Library, Project Gutenberg, Internet Archive, Encyclopaedia Britannica, Old Bailey, Hansard and more, 1800–1899 |
| Pretraining tokens | ~2.9B | 7B to 54B, Chinchilla-optimal for each size |
| Sizes | ~340M | 345M, 500M, 1.2B, 2.7B |
| Tokenizer | nanochat's BPE | 65,536-token byte-level BPE trained on the new corpus, with year, source and OCR-quality conditioning tags |
| Chat fine-tuning | conversations with synthetic, model-written Victorian replies | every reply is genuine 19th-century text |
Models:
On the project's 1,356-prompt behavioural evaluation, Mr.Chatterbox-2.7B-Chat scores 0.741 against 0.655 for this model. That evaluation is scored by an LLM judge on responsiveness, behavioural fit and period register, and does not check factual accuracy.
This model remains available for reference and comparison.
Newer models are available. This is the original Mr. Chatterbox, released in March 2026. It has been succeeded by a new family of Mr. Chatterbox models trained from scratch on a much larger corpus. See Successor models below.
Mr. Chatterbox is a language model 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 (https://huggingface.co/datasets/TheBritishLibrary/blbooks). The model has absolutely no pre-training inputs from after 1899 — the vocabulary and ideas are formed from its training set of nineteenth-century literature.
Mr. Chatterbox's training corpus was 28,035 books, with an estimated 2.93 billion input tokens after filtering. The model has roughly 340 million parameters, roughly the same size as GPT-2-Medium. The difference is, of course, that unlike GPT-2, Mr. Chatterbox is pretrained entirely on historical data.
The model was trained using Andrej Karpathy's Nanochat, with two epochs of supervised fine-tuning and one small supervised fine-tuning epoch for help handling edge cases.
Mr. Chatterbox is a perfect Victorian Gentleman, he's just, you know, a chatbot. Don't tell him that though!
The Mr. Chatterbox project has since trained a new family of models. Everything about the recipe was rebuilt:
| this model | the new family | |
|---|---|---|
| Corpus | 28,035 British Library books, 1837–1899 | ~84.5 GB from HathiTrust, Institutional Books, British Library, Project Gutenberg, Internet Archive, Encyclopaedia Britannica, Old Bailey, Hansard and more, 1800–1899 |
| Pretraining tokens | ~2.9B | 7B to 54B, Chinchilla-optimal for each size |
| Sizes | ~340M | 345M, 500M, 1.2B, 2.7B |
| Tokenizer | nanochat's BPE | 65,536-token byte-level BPE trained on the new corpus, with year, source and OCR-quality conditioning tags |
| Chat fine-tuning | conversations with synthetic, model-written Victorian replies | every reply is genuine 19th-century text |
Models:
On the project's 1,356-prompt behavioural evaluation, Mr.Chatterbox-2.7B-Chat scores 0.741 against 0.655 for this model. That evaluation is scored by an LLM judge on responsiveness, behavioural fit and period register, and does not check factual accuracy.
This model remains available for reference and comparison.
