Uploaded October 2026 | Updated November 2026, 5 days ago
Fine-tune Gemma 3 270M on a MacBook with Unsloth Studio, with no training code: sort support tickets into six queues. 25.3 % → 99.4 % of 1,170 test tickets routed correctly after 15.2 seconds of LoRA training on the Mac's GPU (MLX), then export to GGUF and run it in Ollama (98.9 %). Lesson 4 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, GGUF Q4_K_M): huggingface.co/kgptalkie/gemma-3-270m-ticket-router-GGUF
Run it: ollama run hf.co/kgptalkie/gemma-3-270m-ticket-router-GGUF:Q4_K_M
The dataset (Apache-2.0, synthetic tickets): huggingface.co/datasets/cngchis/Support-Ticket-Router-12K-Cleaned
This is a first version trained for 15 seconds on a small subset of the training tickets, made to show the technique; better versions will be pushed to the same repo.
00:00 The result
00:14 After training: 25.3 → 99.4 %
00:39 The job: six ticket queues
00:54 Gemma 3 270M
01:12 The dataset: 9,354 tickets
01:39 Install Unsloth Studio
02:01 Test notebook
02:44 Base model: 25.3 %
03:08 Studio: model and LoRA
03:42 Dataset and column mapping
04:04 Training parameters
04:50 Start training
05:23 Done in 15.2 seconds, compare in chat
05:56 Tuned model in the notebook
06:18 Score: 99.4 %
06:45 An honest note on scope
07:06 Export to GGUF (Q4_K_M)
07:26 Run it in Ollama
07:45 The model on Hugging Face
07:59 The course
08:17 Recap and next lesson
Lesson 3 (System 1 decision model with Qwen 3.5): youtu.be/wr3dMotYFMU
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/google/gemma-3-270m-it (Gemma terms; Studio uses the identical open copy huggingface.co/unsloth/gemma-3-270m-it)
Unsloth Studio: unsloth.ai
Machine: MacBook Pro, Apple M5 Max, 36 GB unified memory. LoRA 16-bit (≈ 30M trainable parameters), 300 steps × 8 tickets, 15.2 s, 1.47 GB peak memory; exported Q4_K_M file 253 MB.
Fine-tune Gemma 3 270M on a MacBook with Unsloth Studio, with no training code: sort support tickets into six queues. 25.3 % → 99.4 % of 1,170 test tickets routed correctly after 15.2 seconds of LoRA training on the Mac's GPU (MLX), then export to GGUF and run it in Ollama (98.9 %). Lesson 4 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, GGUF Q4_K_M): huggingface.co/kgptalkie/gemma-3-270m-ticket-router-GGUF
Run it: ollama run hf.co/kgptalkie/gemma-3-270m-ticket-router-GGUF:Q4_K_M
The dataset (Apache-2.0, synthetic tickets): huggingface.co/datasets/cngchis/Support-Ticket-Router-12K-Cleaned
This is a first version trained for 15 seconds on a small subset of the training tickets, made to show the technique; better versions will be pushed to the same repo.
00:00 The result
00:14 After training: 25.3 → 99.4 %
00:39 The job: six ticket queues
00:54 Gemma 3 270M
01:12 The dataset: 9,354 tickets
01:39 Install Unsloth Studio
02:01 Test notebook
02:44 Base model: 25.3 %
03:08 Studio: model and LoRA
03:42 Dataset and column mapping
04:04 Training parameters
04:50 Start training
05:23 Done in 15.2 seconds, compare in chat
05:56 Tuned model in the notebook
06:18 Score: 99.4 %
06:45 An honest note on scope
07:06 Export to GGUF (Q4_K_M)
07:26 Run it in Ollama
07:45 The model on Hugging Face
07:59 The course
08:17 Recap and next lesson
Lesson 3 (System 1 decision model with Qwen 3.5): youtu.be/wr3dMotYFMU
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/google/gemma-3-270m-it (Gemma terms; Studio uses the identical open copy huggingface.co/unsloth/gemma-3-270m-it)
Unsloth Studio: unsloth.ai
Machine: MacBook Pro, Apple M5 Max, 36 GB unified memory. LoRA 16-bit (≈ 30M trainable parameters), 300 steps × 8 tickets, 15.2 s, 1.47 GB peak memory; exported Q4_K_M file 253 MB.










