Deploying SLMs Through Conversation

Distil Labs has released a Claude Code skill that lets a single conversation take you from cleaning up production traces to deploying a fine-tuned SLM.

The core flow has 5 steps:

• Generate traces.jsonl (OpenAI message format), job_description.json, and config.yaml
• Filter out low-quality traces and relabel using a committee of teachers
• First verify that the teacher can actually solve the task
• The teacher generates synthetic data and the student trains on it
• After comparing performance, deploy to a managed endpoint or download the weights to self-host

Since training costs consume credits, there are approval checkpoints at each costly stage, where Claude pauses for confirmation.

As a real-world example, 327 noisy production traces from a restaurant-reservation agent were cleaned up and used to fine-tune a Qwen3-1.7B-based multi-turn tool-calling model.

Installation involves running distil signup and then adding distil-cli-skill from the marketplace, and the platform is free to use for open-source projects.