⚡ TLDRs on Google’s Gemma, spoiler: don’t use Gemma 7B (yet?)

👞 Non technical TLDR

• Google “open weighted” 2 models: Gemma-2B and Gemma-7B, which are direct competitors of Microsoft’s phi2 and Mistral AI’s Mistral-7B resp.

• They report that their 7B model is better on most benchmarks than Mistral-7B (the current leader for this “size class”)

• The community (and I) see the contrary so far, but it wouldn’t be the first time that errors in implementations hinder performances, it has only been 4 days, let’s give it some time

• Google created a Gemma tuning competition on Kaggle, awarding $10k / winner: they are pushing adoption

• If Gemma is indeed worse, it could just be seen as a move from Google to be able to sit at the “LLM open [[sourcers]]” table without contributing significant value

🔬 Technical TLDR

• Google released weights for Gemma-2B & Gemma-7B + a technical report + a standalone C++ inference engine implementation

• 2B trained on 2T tokens, 7B trained on 6T tokens. 6T is a lot but rumours say Mistral-7B was trained on 8T

• Both models are now available in llama.cpp and ollama

• Vocabulary 8x larger than Llama2’s (256k vs 32k)

• Google observes better performance with Gemma 7B compared to Mistral 7B, the community (me included) observes the contrary so far

• I tested it using llama.cpp, it would not be the first time that errors in implementation degrades performances, it’s maybe too early to judge, let’s see

• Google introduced yet a new chat template (check attached screenshot)

• They kinda lie on the number of parameters: Mistral-7B has 7.2B parameters; Gemma-7B has 8.5B

• You can try Gemma on Hugging Face Chat & Perplexity: link in the comments

• Waiting for chatbot areana benchmarks

Links in the comments