Uploaded October 2026 | Updated November 2026, 3 days ago
Turn a chat model into a System 1 decision model: fine-tune Qwen 3.5 0.8B on a MacBook with Hugging Face Transformers, PEFT (LoRA) and PyTorch MPS. Like Jev (TypeSafe) and Cloudflare's Clef, it reads a typed request once and returns a typed answer with a probability for every allowed option, in one forward pass, with zero output tokens. Routing 826 real consumer complaints to 9 financial products: 43.6 % as a chat model → 73.1 % after 15 min 36 s of training on the Mac's GPU, no NVIDIA GPU. Lesson 3 of the series Fine-tune LLMs on your Mac.
Full series playlist: youtube.com/playlist?list=PLbDRDyy1iJ_c
My full course, Fine Tuning LLM with Hugging Face Transformers for NLP: kgptalkie.com/fine-tuning-llm
All my courses: kgptalkie.com/udemy-courses
The model (public): huggingface.co/kgptalkie/mini-clef
The dataset: huggingface.co/datasets/kgptalkie/cfpb-complaint-routing
mini-clef is inspired by Cloudflare's Clef (huggingface.co/Cloudflare/clef); it is not made by or affiliated with Cloudflare. It is a quick test run on a small subset (800 complaints), made to show the technique; better versions will be pushed to the same repo.
00:00 The result
00:21 System 2 vs System 1
00:47 Jev and Clef
01:26 What we'll build: mini-clef
01:40 Project and machine
02:34 The data: 826 complaints, 9 products
03:38 The request
03:57 Base model as System 2: 43.6 %
05:10 Clef's prompt layout
05:38 One forward pass
06:06 The joint schema head
07:01 LoRA + the head
07:45 Training on the Mac GPU
08:36 The tuned model
08:59 Before vs after: 43.6 → 73.1
09:36 An honest note on scope
10:18 Save and push to Hugging Face
10:43 The model on Hugging Face
11:17 Download and use it
11:57 Recap and next lesson
Lesson 2 (invoice PDF to JSON with Granite 4.0): youtu.be/1U3u7kqz4zQ
Lesson 1 (text to SQL with MLX): youtube.com/watch?v=tsShl5GSGLE
Base model: huggingface.co/Qwen/Qwen3.5-0.8B
Jev docs: docs.typesafe.ai
Machine: MacBook Pro, Apple M5 Max, 36 GB unified memory. LoRA rank 16 + a 15.8M-parameter joint schema head, 100 steps × 8 complaints, 15 min 36 s, 7.1 GB GPU memory, ≈ 350 ms per complaint.
Turn a chat model into a System 1 decision model: fine-tune Qwen 3.5 0.8B on a MacBook with Hugging Face Transformers, PEFT (LoRA) and PyTorch MPS. Like Jev (TypeSafe) and Cloudflare's Clef, it reads a typed request once and returns a typed answer with a probability for every allowed option, in one forward pass, with zero output tokens. Routing 826 real consumer complaints to 9 financial products: 43.6 % as a chat model → 73.1 % after 15 min 36 s of training on the Mac's GPU, no NVIDIA GPU. Lesson 3 of the series Fine-tune LLMs on your Mac.
Full series playlist: youtube.com/playlist?list=PLbDRDyy1iJ_c
My full course, Fine Tuning LLM with Hugging Face Transformers for NLP: kgptalkie.com/fine-tuning-llm
All my courses: kgptalkie.com/udemy-courses
The model (public): huggingface.co/kgptalkie/mini-clef
The dataset: huggingface.co/datasets/kgptalkie/cfpb-complaint-routing
mini-clef is inspired by Cloudflare's Clef (huggingface.co/Cloudflare/clef); it is not made by or affiliated with Cloudflare. It is a quick test run on a small subset (800 complaints), made to show the technique; better versions will be pushed to the same repo.
00:00 The result
00:21 System 2 vs System 1
00:47 Jev and Clef
01:26 What we'll build: mini-clef
01:40 Project and machine
02:34 The data: 826 complaints, 9 products
03:38 The request
03:57 Base model as System 2: 43.6 %
05:10 Clef's prompt layout
05:38 One forward pass
06:06 The joint schema head
07:01 LoRA + the head
07:45 Training on the Mac GPU
08:36 The tuned model
08:59 Before vs after: 43.6 → 73.1
09:36 An honest note on scope
10:18 Save and push to Hugging Face
10:43 The model on Hugging Face
11:17 Download and use it
11:57 Recap and next lesson
Lesson 2 (invoice PDF to JSON with Granite 4.0): youtu.be/1U3u7kqz4zQ
Lesson 1 (text to SQL with MLX): youtube.com/watch?v=tsShl5GSGLE
Base model: huggingface.co/Qwen/Qwen3.5-0.8B
Jev docs: docs.typesafe.ai
Machine: MacBook Pro, Apple M5 Max, 36 GB unified memory. LoRA rank 16 + a 15.8M-parameter joint schema head, 100 steps × 8 complaints, 15 min 36 s, 7.1 GB GPU memory, ≈ 350 ms per complaint.








