Atlas

Models

← All models

Gemma 7B IT

Google DeepMind · 2024-02-21 · 8.5B parameters

Gemma 7B IT is the instruction-tuned checkpoint of Google's original 7B-class Gemma model. It adapts Gemma 7B for dialogue, instruction following, and assistant-style interaction under the Gemma open-weight terms.

Benchmark scores

BenchmarkScore
AlpacaEval 2.0 (length-controlled win rate)10.4
AlpacaEval 2.0 (raw win rate)6.9
Arena Text — Business, Management & Financial Ops (No Style Control)1025
Arena Text — Chinese (No Style Control)1035
Arena Text — Creative Writing (No Style Control)1024
Arena Text — English (No Style Control)1090
Arena Text — Entertainment, Sports & Media (No Style Control)1002
Arena Text — Exclude Ties (No Style Control)864
Arena Text — Exclude Ties (Style Controlled)975
Arena Text — Expert (No Style Control)1001
Arena Text — French (No Style Control)1024
Arena Text — French (Style Controlled)1127
Arena Text — Hard Prompts (No Style Control)1042
Arena Text — Hard Prompts English (No Style Control)1062
Arena Text — Instruction Following (No Style Control)1018
Arena Text — Legal & Government (No Style Control)1044
Arena Text — Life, Physical & Social Science (No Style Control)1021
Arena Text — Longer Query (No Style Control)1022
Arena Text — Medicine & Healthcare (No Style Control)1027
Arena Text — Multi-turn (No Style Control)963
Arena Text — Non-English (No Style Control)998
Arena Text — Non-English (Style Controlled)1082
Arena Text — Occupational Mathematical (No Style Control)1061
Arena Text — Russian (No Style Control)992
Arena Text — Software & IT Services (No Style Control)1061
Arena Text — Writing, Literature & Language (No Style Control)1034
BBH (Open LLM Leaderboard v2)36.5
Capability96.4
Creative Writing v246.2
EQ-Bench v261.7
Loading Atlas data…