Atlas

Models

← All models

Phi-3 Medium 4K Instruct

Microsoft Research · 2024-05-21 · 14B parameters

Phi-3 Medium 4K Instruct is Microsoft's dense 14B-class instruction-tuned language model with a 4,096-token context window and configured 2,047-token sliding-window grouped-query attention. The released checkpoint was post-trained with supervised fine-tuning and direct preference optimization for instruction following and safety; Microsoft reports an October 2023 cutoff for its offline training data. It is the short-context counterpart to Phi-3 Medium 128K Instruct.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)1098
Arena Text — Chinese (No Style Control)1107
Arena Text — Creative Writing (No Style Control)1107
Arena Text — English (No Style Control)1169
Arena Text — Entertainment, Sports & Media (No Style Control)1082
Arena Text — Exclude Ties (No Style Control)979
Arena Text — Exclude Ties (Style Controlled)1064
Arena Text — Expert (No Style Control)1107
Arena Text — French (No Style Control)1109
Arena Text — French (Style Controlled)1183
Arena Text — German (No Style Control)1101
Arena Text — Hard Prompts (No Style Control)1127
Arena Text — Hard Prompts English (No Style Control)1144
Arena Text — Instruction Following (No Style Control)1114
Arena Text — Japanese (No Style Control)1043
Arena Text — Korean (No Style Control)954
Arena Text — Legal & Government (No Style Control)1131
Arena Text — Life, Physical & Social Science (No Style Control)1138
Arena Text — Longer Query (No Style Control)1120
Arena Text — Medicine & Healthcare (No Style Control)1118
Arena Text — Multi-turn (No Style Control)1088
Arena Text — Non-English (No Style Control)1094
Arena Text — Non-English (Style Controlled)1152
Arena Text — Occupational Mathematical (No Style Control)1176
Arena Text — Russian (No Style Control)1144
Arena Text — Software & IT Services (No Style Control)1134
Arena Text — Spanish (No Style Control)1094
Arena Text — Writing, Literature & Language (No Style Control)1126
BBH (Open LLM Leaderboard v2)64.1
Capability117.3
Loading Atlas data…