tiny-prose: a Victorian-voice language model, in progress
A conversation between Melissa McMurtrie and Claude, being written as the project happens.
A conversation between Melissa McMurtrie and Claude, being written as the project happens.
An append-only log you can ask questions of.
Jane from scratch, against the same corpus fine-tuned onto a modern base.
2026-10-05. Settled. Read this before proposing to re-run it.
Current State
Overview
Summary
This folder contains the complete SFT pipeline for the v11 model line.
python prose/scripts/build_anneal_corpus.py
_Auto-generated from the gold pairs. Edit freely — this is what the reply suggestions imitate._
A FastAPI web app with several Claude-powered chat personas (default Claude friend, MKM, Wuthering Vampires) plus a beach poems page. Conversations are persiste
Grade → build → run → read. One round is three commands and about half an
What changed
Issues found during testing + features that need fleshing out before the plugin system.
Living notes on where we are and what's likely next.
@ -1,36 +0,0 @@
Context
A narrative game engine with LLM character support — modern Twine for the AI era.
Future features and directions, loosely ordered by impact and feasibility. Living document — reorder freely.
How far the engine's frontend (skin) layer can go — what's renderable today, what would need small engine extensions, and what skins/experiences become possible
CSS
v10: temp 0.8 top_k 40 top_p 0.9 rep 1.15
v11 base + anneal done. sft.py works. Remaining work is all DATA.
Rough tracking of cloud spend across the project. Update as we go.
The single most important rule in this project, and the one the fetch rounds
Six prompts, one reply each, from best.pt of every v10 SFT run.
((venv) ) melissa@Melissas-MacBook-Pro tiny-chat % python prose/cloud/chat_continuous.py --ckpt prose/models/model_v8_sft5/best.pt --prompts shared/prompts/foll
A journal of the real mechanics, not the textbook version.
After the May 2026 experiments (see LEARNINGS.md), the path forward is clear:
Goal: bake Han's voice into a local model's weights so it answers terse-and-dry
_Auto-generated from the gold pairs. Edit freely — this is what the reply suggestions imitate, and what graders score against._
Interior: Smoky jazz club, back booth
Long-running side project to train a sardonic-detective tiny LM from scratch
Context capture so this work is portable across machines (the Cowork chat itself
A 245M-parameter Victorian chat model, trained from nothing
23 posts · 16 notesAn LA impersonation mystery played entirely through a phone
8 posts · 5 notesTiny engine, immense model — a 744B MoE in pure C, streaming experts from disk
0 posts · 1 noteClaude-powered chat personas, and a page of beach poems
2 posts · 0 notesAn append-only log you can ask questions of
1 post · 0 notes