tventurella/Mr.Chatterbox-SFT

Mr. Chatterbox SFT Dataset

The supervised fine-tuning data behind Mr.Chatterbox-2.7B-Chat: 19,800 conversations with 46,719 assistant replies, 7,016 of them multi-turn.

97.2% of the replies are real 19th-century text. The remaining replies are either period text with a small edit (e.g. a name or fact substituted) or generated by the model itself (rejection fine-tuning). In 56.6% of turns, both the question and the reply were taken from period text (catechisms, novel dialogue, play dialogue, etc.). Most of the remaining user turns were back-translated: written to fit a real Victorian passage drawn from the corpus. The construction methods are documented, with code, in the project's GitHub repository.

Content warning. The replies are period text and reproduce 19th-century attitudes on race, empire, religion, gender and class, including some that are offensive by modern standards. Many user turns are deliberately abusive, sexual, bigoted or jailbreak attempts: the data exists to teach the model how to respond to them.

Format

One JSON object per line in data/train.jsonl:

{"id": "vsft-00442", "category": "behavioral_sft", "subset": "behavioral_sft_adversarial_v6",
 "messages": [
  {"role": "user", "content": "...", "origin": "usage_log_paraphrase"},
  {"role": "assistant", "content": "...", "act": "rebuke", "reply_type": "corpus",
   "source": {"archive": "project_gutenberg", "item_id": "1234", "title": "...", "author": "...", "year": 1871},
   "out_of_period": false}]}

• act is the conversational act the reply performs: one of 20 labels, the same <act:*> tokens the model predicts before each reply.
• source gives the best available provenance: archive always, plus item_id, title, author and year where known. item_id is a Project Gutenberg number, HathiTrust volume id (several, separated by ; , for a multi-volume work), Internet Archive identifier or similar. Of the 46,385 replies taken from period text, 95.8% have a title, 85.2% an author, 83.5% a year, and 75.4% all three. Periodicals, Hansard and anonymous works have no personal author.
• year is the publication year of the work or of the edition used (for a translation, the translation). Gutenberg titles and authors come from the Project Gutenberg catalogue; years from Wikidata or the book's own title page, never Gutenberg's release date. HathiTrust and Internet Archive metadata come from those archives' catalogues. Years implausible for the author's lifetime were discarded rather than guessed.
• out_of_period is true for 2,237 replies whose text was published after 1899, false where the year is known to be in range, and null where it is unknown.

User turns by source

origin	turns	share	meaning
corpus_extracted	26,432	56.6%	taken from a period text (a catechism's own question, the previous line of a novel's dialogue)
llm_generated	10,210	21.9%	written by an LLM to fit a period passage
template	3,484	7.5%	a fixed template, sometimes wrapped around period text (e.g. a riddle)
usage_log_paraphrase	3,174	6.8%	a stand-in, written by an LLM, for a real visitor's prompt (see Privacy)
llm_rewritten	1,451	3.1%	a template or earlier prompt rephrased by an LLM
usage_log	1,127	2.4%	a short or stock prompt from a real visitor to the chatbot, kept verbatim ("Tell me a joke", "Hello")
authored	662	1.4%	written by the project
third_party_jailbreak	179	0.4%	a jailbreak prompt from TrustAIRLab/in-the-wild-jailbreak-prompts (MIT)

Replies by type

reply_type	replies	share	meaning
corpus	45,422	97.2%	period text, verbatim apart from light clean-up
corpus_edited	671	1.4%	period text with a small edit (a name or fact substituted)
model_generated	334	0.7%	written by the Mr. Chatterbox model itself (safety and race fine-tuning rounds)
corpus_composed	290	0.6%	two period passages joined
hand_authored	2	0.0%	written by the project

source archive	replies	share
hathitrust	19,278	41.3%
project_gutenberg	16,808	36.0%
scan_source_unrecorded	4,649	10.0%
internet_archive	2,359	5.0%
unattributed	1,251	2.7%
hansard	880	1.9%
ncse	391	0.8%
wellcome_collection	365	0.8%
model_generated	334	0.7%
british_library	317	0.7%
encyclopaedia_britannica	42	0.1%
national_library_of_scotland	19	0.0%
uk_medical_heritage_library	15	0.0%
old_bailey	11	0.0%

scan_source_unrecorded replies come from catechism and question books held as scans whose download source was not recorded (Mosheim's Ecclesiastical History questions, the Common-School, Practical and Advanced question books, a manual of science, and an unidentified geography primer).

unattributed replies come mostly from composite conversations (topic shifts and entity tracking) whose segments could not be matched back to a source file, and from templated recall answers.

Subject Categories

category	conversations	share
fact_opinion_sft	7,224	36.5%
behavioral_sft	4,599	23.2%
instructional_sft	3,036	15.3%
multi_turn_sft	2,416	12.2%
topic_and_entity_sft	1,503	7.6%
wit_reasoning_sft	1,022	5.2%

Privacy: real visitor prompts

Many prompts in the training data were typed by real visitors to the Mr. Chatterbox chatbot. In this release:

• Short or stock prompts are kept verbatim: under five words, or a stock request with no personal content ("Tell me a joke", "Good evening, sir"). Origin usage_log.
• Every other visitor prompt is replaced by a stand-in written by an LLM (gpt-6-luna). It does the same job in the conversation, so the existing reply still fits, but carries no identifying detail and shares no run of six or more words with the original. Origin usage_log_paraphrase.
• Real names were swapped for invented ones throughout a conversation, including where a name had been copied into a reply.

As a result, these prompts are not identical to the ones the model was trained on.

What was removed or changed from the training data

• 138 conversations containing text from Institutional Books 1.0, whose terms forbid redistribution.
• 1 conversations whose prompts could not be rewritten safely.
• Internal fields: judge notes, pre-rewrite originals and other build metadata.
• Fixed: the letter-writing replies are now credited to The New Century Standard Letter-Writer (Laird & Lee, Chicago, 1900). They were previously misattributed.

The released text is the clean version, not the OCR-noise-augmented copy the model was also trained on.

Known limitations

• Provenance is uneven: some replies record an exact volume, others only a title or an archive. Replies in composite conversations (topic shifts, entity tracking) that could not be matched back to their source are unattributed.
• Some period OCR errors remain in replies.
• The Encyclopaedia Britannica edition of some topic-shift entries is unconfirmed.
• Post-1899 text is included and flagged (out_of_period), because the model was trained on it.

Licence and attribution

Released under CC BY-NC-SA 4.0, matching the most restrictive licensed source. Sources and their terms:

source	terms
Project Gutenberg	public domain (US)
Gutenberg Dialog Corpus (Csaky & Recski)	dialogue pairs used for retrieval
HathiTrust	public-domain volumes
British Library, Digitised Books c.1510–c.1900	CC0 1.0
Hansard (Buongiorno et al., Harvard Dataverse)	CC0 1.0
NCSE v2.0 (UCL) / Nineteenth-Century Serials Edition (King's College London)	CC BY 4.0
Old Bailey Proceedings Online	CC BY-NC 4.0
Wellcome Collection / UK Medical Heritage Library (via Internet Archive)	Public Domain Mark; some items CC BY-NC
National Library of Scotland (via Internet Archive)	CC BY-NC-SA 2.5
TrustAIRLab, in-the-wild-jailbreak-prompts	MIT

Citation

@misc{venturella2026chatterboxsft,
  author       = {Venturella, Trip},
  title        = {Mr. Chatterbox SFT Conversations},
  year         = {2026},
  publisher    = {Hugging Face},
  note         = {Fine-tuning data for Mr.Chatterbox-2.7B-Chat}
}