DEDUPEDWho it is forAn ML-curious developer with labeled tickets or complaints whose zero-shot accuracy is too low.
The problemGeneric models do not know your routing categories or escalation rules.
The outcomeA fine-tuned Kev with a temperature fitted to your labels, served from your own endpoint.
What you need first
- Python with uv; install from source (pip install kev is a different package)
- A GPU (CUDA) or an Apple Silicon Mac with MLX; Kev-0.8B runs on a laptop
- Optional: a Modal account for fine-tuning and hosted endpoints
Steps
- Run
npx skills add jaredpalmer/kev@kev-finetune - Ask your agent to fine-tune Kev on your support tickets
- Let the skill interview you and find the questions your code already asks
- Train on Modal, check accuracy on held-out data and size your dataset first
- Deploy the result as your own HTTPS endpoint that scales to zero
Prompt for your agent
Use the kev-finetune skill from jaredpalmer/kev to fine-tune Kev-4B on my labeled tickets in <file>. Report accuracy before and after on a held-out split, and stop before deploying if the gain is within noise.
Caveat. The README notes that with about 400 records the gain was inside the noise, so size your dataset before training.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for jaredpalmer/kev and "Fine-tune Kev on your own support tickets with an agent skill"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer