Human Preferences

A local-first tool for capturing your personal taste as a dataset.

Judge thousands of A/B pairs → export a profile that teaches any model to judge the way you would, or to fine-tune a proxy of you.

---

Why

Two goals, one schema:

1. A taste profile / LLM-judge rubric — attach your revealed preferences to an existing model so it can rank subjective work (screenplays, designs, copy…) the way you would.
2. A mindfile — accumulate enough personal preference data to one day fine-tune a proxy of your judgment.

The data shape mirrors Stanford Human Preferences (SHP) and maps cleanly to DPO/RLHF — but the content is yours, not the crowd's.

Quickstart

npm install --cache ./.npm-cache    # see note on the npm cache below
cp .env.example .env                 # edit USER_DISPLAY_NAME etc.
npm run db:migrate                   # create data/preferences.db
npm run seed:demo                    # 46 curated pairs to start judging now
npm run dev                          # → app on http://localhost:5173, API on :8787

Open the app and start judging. Keyboard: A / B to choose, N for no preference, S to skip. Add an optional note on why — it's gold for the rubric.

npm cache: if npm install fails with EACCES on ~/.npm, this repo already points npm at a project-local cache via .npmrc. Run commands under caffeinate -i on macOS so long batches don't get interrupted by sleep.

How it works

 React SPA (Vite, :5173) ──/api──▶ Fastify (:8787) ──▶ SQLite (WAL)
        the judging UI                  REST              one file = your dataset

• A pair is the SHP/DPO triple: context + option_a + option_b.
• A judgment is your choice (a/b/skip/no_preference) + optional note + latency.
• Everything carries schema_version; dedup is a content hash of the normalized, order-independent context|A|B. Schema lives in schema/ — the SQL migration and a portable preference-record.v1.json.

Filling the queue (generation)

Generation never calls a model from inside the app — pairs arrive via a simple file/ingest protocol, from three sources:

Source	How
Local Gemma (bulk)	export-job writes a job to a synced folder; a Python runner (gemma-runner/) on a second machine produces pairs with Ollama; ingest-batch reads them back.
SHP corpus	npm run import:shp -- --spread --limit 2000 streams Reddit posts as seed topics (not pre-made pairs); generation turns each into opinionated A/B stances.
Claude Code	Drive Claude Code in this repo to author high-quality or current-events pairs directly, then ingest-file.

Plus two non-text regions: Narrative ("which would you rather watch/read?" across film/tv/books/music/games) and Visual (npm run fetch:images pulls a CC0 art set from the Art Institute of Chicago into the repo; npm run import:images pairs them — the Judge view renders images in the same A/B chrome).

See CLAUDE.md for the exact commands and the GeneratedPairLine shape.

Progression — the Taste Map

The dashboard isn't a flat queue. Your themes group into regions (ideas / daily life / power / culture / narrative / visual / the now), and a single Robustness score (breadth × depth × consistency) tracks how complete your profile is. Judge in sets of ~20 that span regions; a tier ladder (Sketch → Calibrated → Proxy-ready → Robust → High-fidelity → Mirror → Mindfile) marks what each volume of judgments unlocks — from an LLM-judge rubric to a full fine-tune. All typographic; no charts.

Exporting your taste

From the Profile page, or the CLI:

npm run export:dpo      # exports/dpo.jsonl  → {prompt, chosen, rejected, meta}
npm run export:rubric   # exports/taste-rubric.md → an LLM-judge system prompt

• DPO/RLHF JSONL — drop into a DPOTrainer-style pipeline. Skip / no-preference are excluded; provenance is preserved in meta.
• Taste rubric — a system prompt synthesized from your real choices (with your notes), grouped by theme and weighted by sample size. Paste it into any model so it judges like you.

Privacy

Your preferences are yours. The code is open-source; your data is not. .gitignore keeps data/ (the SQLite DB), exports/, and .env out of git from the first commit. Back up your dataset by copying data/preferences.db.

Project layout

schema/      SQL migrations + portable export JSON Schema
shared/      domain + protocol types (one source of truth)
server/      Fastify API + better-sqlite3
app/         Vite + React SPA (editorial UI)
worker/      generation exchange CLI (export-job / ingest-batch / import:shp)
export/      DPO + rubric builders
gemma-runner/  Python batch runner for local Ollama/Gemma (runs on a 2nd machine)

License

MIT — see LICENSE.