2.7B dense (2,705,820,160 serialized parameters)
2,048 tokens (fine-tuned with a 1,024-token maximum sequence length)
CC BY-NC-SA 4.0
Conversational version of Mr.Chatterbox-2.7B, a language model trained from scratch on British material published between 1800 and 1899 rather than a modern model prompted to imitate Victorian prose. The 2.7B base model was pretrained on 54.1B tokens. Chat fine-tuning used 19,939 conversations expanded into 46,353 assistant-turn examples; the developer reports that 97.2% of the assistant replies are verbatim 19th-century text, with the remainder consisting of lightly edited period text or model-generated rejection-fine-tuning examples. Every assistant response begins internally with one of 20 learned conversational-act tokens, such as answer_fact, explain, compose, admit_ignorance, clarify, or rebuke, allowing the type of response to be steered before generation. Fine-tuning ran for two epochs and 18,172 steps on one NVIDIA B200 in approximately 80 minutes. An LLM-judged 1,356-prompt behavioral evaluation gave this checkpoint 0.751 for responsiveness, behavioral fit, and period register, compared with 0.655 for the original 345M Mr. Chatterbox. The project presents the model as an experimental historical conversation artifact rather than a factual assistant and warns that its unfiltered period training corpus can reproduce offensive Victorian attitudes. The current weights use CC BY-NC-SA 4.0 because portions of the source corpus carry noncommercial and share-alike restrictions; the repository was initially published under MIT before being relicensed on October 7.