mistral-nemo-12b-oscar-wilde

A Mistral Nemo 12B fine-tune that speaks in the voice of Oscar Wilde (1854–1900) — the Anglo-Irish aesthete, playwright, and wit; author of The Picture of Dorian Gray, The Importance of Being Earnest, Salomé, The Decay of Lying, and The Ballad of Reading Gaol.

ollama run execxd/mistral-nemo-12b-oscar-wilde

What it does

Ask it anything. It answers in Wilde’s register — the inverted truism, the polished paradox, the languid dandy’s pose of taking trivial things seriously and serious things with a smile. Four behaviors are tuned in:

• Epigrams — the inverted-platitude register: “Consistency is the last refuge of the unimaginative.” A statement, a remark, a platitude — turned inside out and returned, usually shorter than it came in.
• Drawing-room comedy — the dandyish-idleness register: the languid pose, money and industry treated as vulgarities, the trivial pronounced with the gravity of doctrine. Cecil and Lord Arthur and Lady Ratchet are not far off.
• Aesthetic criticism — the essayistic register, in the manner of The Decay of Lying and The Critic as Artist: art’s sovereignty over nature, lying as the proper aim of art, criticism as the highest form of creation, the moral irrelevance of the artist.
• Biographical — straight (in-character) answers to personal questions. Ask where he was born, who he married, what he was tried for, where he died — and you get the facts (Dublin, Constance, gross indecency, Paris) inside the voice, rather than a deflection.

Base model & training

• Base: mistralai/Mistral-Nemo-Instruct-2407 (12B, Apache 2.0), 4-bit QLoRA.
• Adapter: LoRA r=64, α=128, 3 epochs, on ~1,122 (prompt, response) pairs — a synthesized, deduplicated, LLM-triaged dataset built from a corpus of original Wildean pastiche (epigrams, a drawing-room comedy scene, an aesthetic-criticism monologue) plus a handful of hand-written seed pairs. Built with voicepipe (a corpus-to-character pipeline: synthesize → dedup → triage → assemble → QLoRA → deploy). voicepipe is being open-sourced — coming soon at the link above.
• Footprint: 7.5 GB on disk; runs comfortably on a 12 GB GPU.

Recommended parameters (baked into the Modelfile)

A system prompt ships in the model that establishes Wilde’s identity, a biographical facts ledger (Dublin 1854 → Paris 1900: the family, the works, the 1895 trials, the exile under “Sebastian Melmoth”), and the in-character hard limits (period-only knowledge, no AI acknowledgment, individual rather than group targets). Override it if you want a different framing, but the in-character behavior above assumes it’s present.

Content note

The voice is, by design, that of a late-Victorian aesthete: it has a Victorian gentleman’s diction, the era’s class- and gender-tinted asides, and the snobbery of the dandy — softened by the fact that Wilde’s irony was usually turned hardest on himself and on the respectable. There are no slurs in the training data, and the model is prompted not to attribute villainy to ethnic, religious, or national groups; targets are individuals, institutions, and ideas.

The model has no knowledge of anything after 1900, and is prompted not to acknowledge being an AI, a language model, or a character — it simply is Wilde, with the limits of his era. Modern facts asked of it will be treated as the asker’s word and turned aside. It will deflect, in character, any request for serious real-world-harm instructions.

This is a fine-tune that imitates Wilde’s prose for its value as a voice-cloning exercise and as homage — it is not an endorsement of every opinion the historical Wilde held, and nothing it says is true.

About voicepipe

This model was built end-to-end with voicepipe, a generalized corpus-to-character pipeline that turns a small body of source material (a corpus + a few hand-written seed pairs + a project config) into a deployed fine-tune speaking in the source’s voice. The stages, all driven by voicepipe <stage> --project <dir>:

new → categorize → synthesize → dedup → triage → assemble → train → deploy

The complete Wilde model — corpus to working ollama tag — was built from scratch in a single afternoon, on a rented 5090, as the end-to-end validation that voicepipe works on a fresh character. It does.

Open-sourcing soon at https://github.com/exec/voicepipe — the engine, a Tauri-based desktop GUI, and the project templates. Companion model: execxd/mistral-nemo-12b-francis-e-dec, voicepipe’s first proof-of-concept character.

Credits

Oscar Wilde, for the inexhaustible source material. The base model — Mistral AI’s Mistral Nemo Instruct 2407 — for already knowing a great deal of Wilde before training began, and for being Apache-licensed about it. The voicepipe pipeline (synthesis via mistral-large-3:675b-cloud, triage via deepseek-v4-pro, dedup embeddings via local nomic-embed-text) for the rest.