Atlas

Models

← All models

Mistral 7B Instruct v0.2

Mistral AI · 2023-12-11 · 7.2B parameters

Mistral 7B Instruct v0.2 is Mistral AI's instruction-fine-tuned version of Mistral 7B v0.2, released on December 11, 2023 with Apache-2.0 open weights. Its v0.2 base raises the context window from 8K to 32K, sets RoPE theta to 1e6, and removes sliding-window attention; it uses Mistral Tokenizer V1.

Benchmark scores

BenchmarkScore
AlpacaEval 2.0 (length-controlled win rate)17.1
AlpacaEval 2.0 (raw win rate)14.7
Arena Text — Business, Management & Financial Ops (No Style Control)1038
Arena Text — Chinese (No Style Control)1008
Arena Text — Creative Writing (No Style Control)1068
Arena Text — English (No Style Control)1134
Arena Text — Entertainment, Sports & Media (No Style Control)1058
Arena Text — Exclude Ties (No Style Control)915
Arena Text — Exclude Ties (Style Controlled)999
Arena Text — Expert (No Style Control)1035
Arena Text — French (No Style Control)1036
Arena Text — French (Style Controlled)1114
Arena Text — German (No Style Control)987
Arena Text — Hard Prompts (No Style Control)1067
Arena Text — Hard Prompts English (No Style Control)1091
Arena Text — Instruction Following (No Style Control)1060
Arena Text — Japanese (No Style Control)879
Arena Text — Legal & Government (No Style Control)1099
Arena Text — Life, Physical & Social Science (No Style Control)1079
Arena Text — Longer Query (No Style Control)1060
Arena Text — Medicine & Healthcare (No Style Control)1067
Arena Text — Multi-turn (No Style Control)1062
Arena Text — Non-English (No Style Control)1012
Arena Text — Non-English (Style Controlled)1077
Arena Text — Occupational Mathematical (No Style Control)1086
Arena Text — Russian (No Style Control)1017
Arena Text — Software & IT Services (No Style Control)1084
Arena Text — Spanish (No Style Control)1025
Arena Text — Writing, Literature & Language (No Style Control)1073
BBH (Open LLM Leaderboard v2)44.6
Loading Atlas data…