Fine-tune Qwen3 with Codex and Hugging Face Jobs
Learn the Codex fine-tuning workflow for Qwen3-0.6B, including setup and a 60-ticket evaluation that reports accuracy rising from 20% to 85%.
Player not loading? Watch on YouTube
This tutorial uses Codex and the Hugging Face Model Trainer skill to fine-tune Qwen/Qwen3-0.6B into a customer-support ticket router. The Bitext dataset supplies customer messages and intent labels. The requested output is JSON with an intent, confidence score and short reason.
The walkthrough covers a UV environment, Hugging Face authentication, skill installation and the Hugging Face MCP server. The presenter lists a Pro account and a token with repository write and job-management permissions as prerequisites. Training runs on Hugging Face Jobs; the proposed self-hosted deployment uses Transformers and a FastAPI application on the user's server.
The training prompt specifies supervised fine-tuning, a maximum sequence length of 512 and the 12 most frequent intents in the first 5,000 training rows. Codex submits the job and provides monitoring information. The presenter reports about 55 minutes of job time and a cost of roughly 36 to 37 cents for this run.
The notebook's 60-row evaluation reports accuracy rising from 20% to 85%, valid JSON from 28% to 98%, and schema compliance from 28% to 88%. These results differ from the job's reported 100% JSON and schema rates. A final order-change example compares the base model's repetitive response with the tuned model's structured answer. The results concern this task and evaluation sample, rather than general model quality.