Summary
- Original article title, translated: "Google ML Bootcamp: Building a Financial Advisor Chatbot with Gemma-2B." Only the filename's unsupported colon was replaced. Published 2024-10-28 · slug
google-ml-bootcamp · category "AI and machine learning studies / Bootcamps and LLMs" · source article recorded in Korean and English. - A retrospective on the Google Machine Learning Bootcamp hosted by Google Developers Group over five months, June through October 2024. An early used-car price prediction competition strengthened the fundamentals with Scikit-learn, XGBoost, LightGBM, Grid Search, and Bayesian Optimization.
- For the final Gemma Sprint, I built a financial advisor chatbot based on
gemma-2-2b-it, personalized to user profiles. I overcame out-of-memory errors on Kaggle T4 hardware using 4-bit NF4 quantization, reducing weight memory by about 75%, and LoRA (r=64, alpha=16, approximately 0.05% trainable parameters). - After three epochs with SFTTrainer, I merged the LoRA weights to produce a standalone model.
- Lessons: data and prompt design mattered more than model size, and LoRA was a practical choice for a personal GPU setup.
Mentioned
- Google ML Bootcamp 2024: the completed bootcamp discussed in the article.
financial_advisor: the Gemma Sprint product, a financial advisor chatbot.- Google: the Gemma model and the organizing Google Developers Group community.
- Kaggle: the training environment, Kaggle Notebook with a T4.
Connections
- Writing · Credentials · Projects
- Atomic notes: Data and prompt design matter more than model size; LoRA is practical on a personal GPU · Q: Is the Google ML Bootcamp financial chatbot the same project as
financial_advisor? - Concepts: 4-bit quantization (NF4) · Gemma-2B · LightGBM · LoRA and PEFT · XGBoost · i18n and multilingual support · Running a technical blog · Supervised fine-tuning (SFT) · Prompt design · Hyperparameter tuning
Local graph
Notes 1 · Links 0