Using and fine-tuning Gemma 3
summary
Gemma 3 ships a Gemma Python library that provides chat, multi‑turn caching, and multimodal prompts with spatial tokens, and it integrates with Kauldron integration to run supervised fine‑tuning using HuggingFace, TFDS or JSON datasets and a transform API for tokenizing and padding. Models auto‑shard across TPUs (Colab TPU v2‑8 demo), single‑GPU training uses a LoRA wrapper with partial‑update optimizers, and the release includes weight surgery utilities to extract or fuse LoRA adapters after training.
AI-written summary · about
chapters (6)
- 0:00Gemma Python library
- 1:50Model training with Kauldron
- 4:42Configuration and CLI
- 6:06Scaling and sharding
- 7:35Low-rank adaptation
- 8:31Resources and future roadmap
related talks
10mDemo: Using Gemma with the Hugging Face ecosystem
05mBuild your own LLM on Cloud
44mBuilding amazing Front-End Project with Gemma: A State-of-the-Art Open Model
39mFlutter Project with Gemma: A State-of-the-Art Open Model
10mDemo: Building a Gemma chatbot with Keras
01h 15mWorkshop: Working with Gemma and Open LLMs on Google Kubernetes...
12mDemo: Building cloud-native, AI-powered applications with GKE
08mDemo: JAX, Flax and Gemma