Open Source Model Training

What Is This Program?

Students move from "API consumers" to "Model Owners." Deep-dive into taking a generic foundation model (Llama 3.1 or Mistral) and fine-tuning it for specialized tasks using No-Code tools.

Who Is It For?

AI enthusiasts, developers, and data hobbyists aged 16+ wanting control, privacy, and ownership.

What Will I Learn?

Fine-tuning with HuggingFace AutoTrain, Cloud Training (MonsterAPI), Local Inference (Ollama), and Training Visualization (Weights & Biases).

Pre-Requisites

Data Literacy (Spreadsheets), Logical Workflow Mapping, Conceptual AI Understanding.

Dates

Session 1: July 6 - July 17, 2026
Session 2: August 3 - August 14, 2026 (EN)

Language

This course is offered in both English or Chinese. Check the dates for your language preference.


Tuition & Fees

Tuition: NT$XX,000
Deposit: NT$2,000
Early Bird Deal: Save 15% (Book by March 1st)

Core Tech Stack

Hugging Face

The hub for hosting, sharing, and downloading open weights models and datasets.

Ollama / vLLM

Tools for serving LLMs locally with optimized inference speeds.

Axolotl / Unsloth

Efficient fine-tuning frameworks (QLoRA) to customize models on consumer hardware.

GGUF / Llama.cpp

Quantization formats for running massive models on standard laptops (CPU/metal).

Curriculum

Week 1: Running & Serving Models

Day Topic Hands-On Activity
01 Model Zoos The Download: Pulling Llama-3-8B from Hugging Face.
02 Local Inference The Localhost: Running models via Ollama and LM Studio.
03 Quantization The Shrink Ray: Converting models to GGUF (4-bit).
04 Prompt Templates Chat Formats: Understanding ChatML and Instruction tuning.
05 API Serving The Endpoint: Exposing a local model as an OpenAI-compatible API.

Week 2: Customization & Fine-Tuning

Day Topic Hands-On Activity
06 Dataset Prep The Clean-Up: Formatting JSONL data for training.
07 LoRA & PEFT Parameter Efficient Tuning: Concepts and adapters.
08 Fine-Tuning Run The Training: Fine-tuning Mistral on a custom dataset using Unsloth.
09 Evaluation The Benchmark: Testing the fine-tuned model vs base model.
10 Merged Models The Frankenstein: Creating a model merge (SLERP/TIES).
Enroll in OSS Track