Atlas

Models

← All models

Gemma 2 9B

Google DeepMind · 2024-06-27 · 9.2B parameters

Gemma 2 9B is Google DeepMind's text-to-text, decoder-only, open-weight pre-trained model in the Gemma 2 family. Trained on 8 trillion primarily English tokens using knowledge distillation from a larger teacher, it alternates 4,096-token local sliding-window attention with 8,192-token global attention.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)53.8
BoolQ85.7
Capability114.5
Epoch Capabilities Index (ECI)119
Global PIQA - Non-Parallel (Log-Likelihood Accuracy)60.5
Global PIQA - Parallel (Log-Likelihood Accuracy)27.3
GPQA (Open LLM Leaderboard v2)32.9
GSM8K84.9
IFEval (Open LLM Leaderboard v2)20.4
MATH Level 5 (Open LLM Leaderboard v2)13.4
MMLU-Pro (Open LLM Leaderboard v2)41.0
MuSR (Open LLM Leaderboard v2)44.6
ObviousBench42.4
PIQA83.7
Loading Atlas data…