A truly multilingual Gemma 3
summary
Gemma 3 claims state‑of‑the‑art multilingual fluency by pretraining on 140 languages in pre-training and providing over 35 languages out of the box after instruction tuning, doubling multilingual data share and using a Gemini tokenizer that favors non‑English text. Internal evals report parity with GPT‑4 on instruction tasks, 80% of users are outside the US, and a Kaggle contest drew 7,000 entries. A regional variant SEA‑LION (Gemma 2 9B) is tailored to Southeast Asian languages, and small instruction sets like as few as 40 samples noticeably improve performance.
AI-written summary · about
chapters (9)
- 0:00Introduction to Gemma
- 0:31Global reach and development
- 1:26Performance benchmarks
- 2:24Technical improvements
- 3:50Multimodal application example
- 4:36Diverse industry use cases
- 5:41Southeast Asian language model
- 6:40Project Aquarium and data
- 7:28Tips for developers and closing
related talks
19mGemini 2.0: Beyond the Basics
54mGemini, Google's Large Language Model
51mGemini, Google's Large Language Model
10mAI for everyone with Gemma
16mOpen models in the Gemini era
10mDemo: Using Gemma with the Hugging Face ecosystem
39mFlutter Project with Gemma: A State-of-the-Art Open Model
51mGemini, Google's Large Language Model