Google releases Gemma 2 2B, ShieldGemma and Gemma Scope

Source: HuggingFace Date: 2024-07-31 URL: https://huggingface.co/blog/gemma-july-update

Summary

Model releases: Three additions to the Gemma family — Gemma 2 2B (2.6B params, on-device targeting), ShieldGemma (safety content moderation classifiers in 2B/9B/27B), and Gemma Scope (sparse autoencoders for interpretability). Gemma 2 2B-it: Open LLM Leaderboard v2 mean 17.0 (vs Phi-2 at 15.5, Qwen 1.5B at 13.9), strong on IFEval (56.7). ShieldGemma 9B on SG Prompt: F1 0.828 / AU-PRC 0.894. Gemma 2 2B also functions as speculative decoder for 27B model with up to 3x speedup.

Implications

Model release cadence. Gemma 2 2B as speculative decoder for the 27B model is a practical deployment optimization — the architecture compatibility within the Gemma family means the small model can generate draft tokens the large model accepts, cutting inference latency without accuracy loss. This makes the 27B model viable for latency-sensitive applications that couldn’t previously run it.

Open-weights ecosystem health. ShieldGemma releasing three safety classifier sizes covering different computational budgets makes content moderation accessible as a modular add-on to any open-weights deployment. Gemma Scope with SAELens integration is the most research-valuable release of the three — sparse autoencoders for Gemma 2 enable mechanistic interpretability work on a widely deployed open model family, not just on GPT-2 or toy models.