Atlas

Models

← All models

Gemma 2B IT

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

Gemma 2B IT is Google's original 2B-class instruction-tuned Gemma checkpoint, a text-to-text decoder-only model released with open weights. It post-trains the 2B base model with supervised fine-tuning and RLHF on English prompt-response and preference data for dialogue and instruction following.

Benchmark scores

BenchmarkScore
AlpacaEval 2.0 (length-controlled win rate)5.4
AlpacaEval 2.0 (raw win rate)3.4
Arena Text — Business, Management & Financial Ops (No Style Control)999
Arena Text — Chinese (No Style Control)985
Arena Text — Creative Writing (No Style Control)987
Arena Text — English (No Style Control)1033
Arena Text — Entertainment, Sports & Media (No Style Control)968
Arena Text — Exclude Ties (No Style Control)780
Arena Text — Exclude Ties (Style Controlled)904
Arena Text — Hard Prompts (No Style Control)989
Arena Text — Hard Prompts English (No Style Control)998
Arena Text — Instruction Following (No Style Control)971
Arena Text — Legal & Government (No Style Control)1020
Arena Text — Life, Physical & Social Science (No Style Control)950
Arena Text — Longer Query (No Style Control)981
Arena Text — Medicine & Healthcare (No Style Control)937
Arena Text — Multi-turn (No Style Control)945
Arena Text — Non-English (No Style Control)958
Arena Text — Non-English (Style Controlled)1052
Arena Text — Occupational Mathematical (No Style Control)1004
Arena Text — Russian (No Style Control)936
Arena Text — Software & IT Services (No Style Control)1021
Arena Text — Writing, Literature & Language (No Style Control)978
BBH (Open LLM Leaderboard v2)31.5
Capability81.5
Creative Writing v241.1
EQ-Bench v223.3
EQ-Bench v2 + MAGI-Hard Combined23.7
GPQA (Open LLM Leaderboard v2)27.9
IFEval (Open LLM Leaderboard v2)26.9
Loading Atlas data…