Atlas

Models

← All models

Gemma 2B

Google DeepMind · 2024-02-21 · 2.5B parameters

Gemma 2B is the base (pretrained, not instruction-tuned) 2B checkpoint of Google's original Gemma family, released alongside Gemma 7B in February 2024. It is a text-to-text, decoder-only Transformer with multi-query attention, an 8,192-token context, and a SentencePiece tokenizer, trained primarily on English web documents, mathematics, and code. At roughly 2.5 billion parameters, it targets CPU, laptop, and other resource-constrained or on-device deployments; the separate gemma-2b-it checkpoint is the instruction-tuned variant.

Benchmark scores

BenchmarkScore
ARC AI242.1
BBH (BIG-Bench Hard)35.2
BBH (Open LLM Leaderboard v2)33.7
BoolQ69.4
Capability79.8
Epoch Capabilities Index (ECI)92.6
GPQA (Open LLM Leaderboard v2)25.5
GSM8K17.7
HellaSwag (Unspecified Scoring Protocol)71.4
IFEval (Open LLM Leaderboard v2)20.4
MATH Level 5 (Open LLM Leaderboard v2)3.0
MedMCQA (Open Medical-LLM)30.2
MedQA 4-options (Open Medical-LLM)29.2
MMLU42.3
MMLU Anatomy (Open Medical-LLM)37.0
MMLU Clinical Knowledge (Open Medical-LLM)35.5
MMLU College Biology (Open Medical-LLM)33.3
MMLU College Medicine (Open Medical-LLM)31.2
MMLU Medical Genetics (Open Medical-LLM)28.0
MMLU Professional Medicine (Open Medical-LLM)18.0
MMLU-Pro (Open LLM Leaderboard v2)13.7
MuSR (Open LLM Leaderboard v2)39.8
PIQA77.3
PubMedQA (Open Medical-LLM)66.4
TriviaQA53.2
WinoGrande65.4
Loading Atlas data…