Atlas

Models

← All models

Zephyr 7B Beta

Hugging Face · 2023-10-26 · 7.2B parameters

Zephyr 7B Beta is the second model in Hugging Face H4’s Zephyr assistant series and a full-parameter fine-tune of Mistral 7B v0.1. The final checkpoint was initialized from one epoch of distilled supervised fine-tuning on the filtered UltraChat corpus, then optimized for three DPO epochs on binary preferences constructed from UltraFeedback. It retains the dense 32-layer Mistral architecture with grouped-query attention, a 4,096-token sliding window, and a 32,768-position configuration; the reported experiments used AdamW with no weight decay and bfloat16 precision.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)1000
Arena Text — Chinese (No Style Control)959
Arena Text — Creative Writing (No Style Control)1103
Arena Text — English (No Style Control)1108
Arena Text — Entertainment, Sports & Media (No Style Control)1063
Arena Text — Exclude Ties (No Style Control)889
Arena Text — Exclude Ties (Style Controlled)973
Arena Text — Expert (No Style Control)1005
Arena Text — French (No Style Control)1030
Arena Text — French (Style Controlled)1098
Arena Text — German (No Style Control)989
Arena Text — Hard Prompts (No Style Control)1027
Arena Text — Hard Prompts English (No Style Control)1050
Arena Text — Instruction Following (No Style Control)1032
Arena Text — Legal & Government (No Style Control)1095
Arena Text — Life, Physical & Social Science (No Style Control)1060
Arena Text — Longer Query (No Style Control)1042
Arena Text — Medicine & Healthcare (No Style Control)1026
Arena Text — Multi-turn (No Style Control)1038
Arena Text — Non-English (No Style Control)987
Arena Text — Non-English (Style Controlled)1048
Arena Text — Occupational Mathematical (No Style Control)1044
Arena Text — Russian (No Style Control)1061
Arena Text — Software & IT Services (No Style Control)1049
Arena Text — Spanish (No Style Control)1073
Arena Text — Writing, Literature & Language (No Style Control)1086
BBH (Open LLM Leaderboard v2)43.2
Capability102.4
EQ-Bench v258.3
EQ-Bench v2 + MAGI-Hard Combined47.1
Loading Atlas data…