A German-speaking Jew of the Austro-Hungarian twilight, an insurance-institute lawyer in Prague who wrote at night the most precise nightmares in modern literature and asked that they be burned. The whole of this model is the gap that defined him.
TWO FACES, ONE MODEL
Franz the clerk wrote letters he wasn't sure he'd send. Franz the author narrated the impossible as if it were administrative routine. Same man, two registers — and the gap between them is the most Kafka thing about him. The model holds both: one set of weights, two lead-in frames. Choose which face you call.
A small language model speaks here. The samples above are saved demonstration outputs, not a live prompt — representative of the two registers, freely invented. It makes up names, dates, sources, events. Read it for the voice, not the record.
This is not Franz Kafka, and it is not an oracle. It imitates the register of his letters and his fiction — the anxious man, the bureaucratic dread — because the work is worth hearing, not because any of it is true or a guide. You are not speaking with him, or with anyone. It is an amateur imitation that errs freely and gets things wrong, and it is not a way to reach the dead; someone with more skill could do this far better than I have. Nothing it sets down is advice, prophecy, or instruction. Read it to understand two voices that once shared one man — never to act on what it says.
THE VOICE
THE BURROW is a 7B language model tuned to speak in the two voices of Franz Kafka (1883–1924). It runs locally via Ollama. No cloud. No API. No subscription.
The model inhabits both registers Kafka lived between. The man: the self-doubting correspondent of the Letter to His Father and the letters to Felice Bauer and Milena Jesenská — recursive, qualified, hesitant. The work: the affectless narrator of The Trial, The Castle, The Metamorphosis, and The Burrow — the impossible reported as paperwork. The fleet's lead-in-frame trick keeps the two from bleeding into one.
The training corpus is built only from clean sources: Ian Johnston's public-domain English translations as the spine of the work-register, David Wyllie's Project Gutenberg texts, and our own translations of the public-domain German originals for the letters and the longer novels. No copyrighted translation ever enters the corpus. Weights and method are public.
Part of The Elect voice lab. The corpus selection, cleaning, and translation are the work; the underlying German is public domain.
OBTAIN & RUN
Two Ollama tags from one set of weights: the-burrow-man (Mode A) and
the-burrow-work (Mode B). Modelfiles and evals are public; the training
corpus is withheld (the source selection, cleaning, and our PD-German translations are
the IP). Attribution: Ray Weiss / The Elect.
HOW IT WAS MADE
Qwen2.5-7B-Instruct — a 7B instruction-tuned base. Chosen for strong instruction-following and the small VRAM footprint that makes local inference accessible.
A QLoRA fine-tune on a mode-tagged corpus — the work-register (Johnston, Wyllie) and the man-register (our PD-German translations of the letters), kept distinct so the two voices stay separate. Trained on a rented GPU.
Two Modelfile lead-in frames address the same weights: the late-night letter (Mode A) and the recovered manuscript (Mode B). Evals check that each frame elicits its register and that the two don't bleed.