# Kev Finetune

Fine-tune a Kev decision model (open Jev-style System One model) on a user's own...

- Repository: https://github.com/jaredpalmer/kev (9751 stars)
- Category: DevOps
- License: Apache-2.0
- Last updated: 2026-10-27
- Runs scripts: yes, review before use

## Install

Claude Code: `npx skills add jaredpalmer/kev --skill kev-finetune -g -a claude-code -y`
Codex: `npx skills add jaredpalmer/kev --skill kev-finetune -g -a codex -y`
Cursor: `npx skills add jaredpalmer/kev --skill kev-finetune -g -a cursor -y`
Gemini CLI: `npx skills add jaredpalmer/kev --skill kev-finetune -g -a gemini-cli -y`

## Description

Fine-tune a Kev decision model (open Jev-style System One model) on a user's own workload and serve it with calibrated probabilities. Interviews the user, finds existing Jev/TypeSafe questions or labelled data in their codebase, generates enough synthetic training data to measure a gain, trains and calibrates on Modal, scores against the released checkpoint, deploys a TypeSafe-compatible endpoint, and tears everything down. Use when someone wants noul/choice/score questions answered on their own domain (routing, triage, moderation, classification, scoring), wants calibrated confidence for automation thresholds, mentions Jev, TypeSafe, System One or Kev, or asks to fine-tune, evaluate, deploy or clean up Kev.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Fine-tune Kev on the user's workload

Jev (TypeSafe's System One model) answers typed questions about a text without generating tokens, but it is a fixed
hosted model: on the user's data it is out of distribution and its probabilities cannot be recalibrated. Kev is the open
reconstruction; because it can be trained, you can fine-tune it on a few hundred to a few thousand labelled examples of
the user's exact questions and fit its temperature on a held-out slice, so the probabilities it serves are calibrated for
*their* data. This skill takes a user from "I have this decision to automate" to a deployed, measured endpoint, then
cleans up. Nothing needs a local GPU or a clone of the Kev repo.

Scripts (`scripts/`) are short, standard-library Python plus one Modal app. If the user's case does not fit them,
read the script and change it; they are meant to be edited, not worked around.

| Script | Purpose |
| --- | --- |
| `extract_workload.py` | scan a codebase for Jev / TypeSafe / System One calls and labelled data; draft `workload.json` |
| `convert_data.py` | CSV / JSONL of existing labelled examples -> Kev records (column mapping) |
| `generate_data.py` | workload spec -> balanced labelled records via any OpenAI-compatible model; `--dry-run` prints the prompt |
| `plan_size.py` | how many records make the fine-tune vs baseline comparison statistically meaningful |
| `split_data.py` | validate records, split by state into train / calibration / development |
| `kev_modal.py` | Modal: `validate`, `train`, `evaluate`, `compare`, `pull`, `publish`, `teardown`, and the `Serve` endpoint |

Run `modal` as `uvx modal ...` if it is not installed. Commands below run from the skill directory.

## Phase 0: interview

Do not generate anything before you can answer these. Ask what you cannot infer; keep it to one round if possible.

The rest: https://github.com/jaredpalmer/kev/blob/HEAD/skills/kev-finetune/SKILL.md