Atlas

Models

← All models

DeepSeek LLM 67B Chat

DeepSeek · 2023-11-29 · 67.4B parameters

DeepSeek LLM 67B Chat is the instruction-tuned checkpoint initialized from DeepSeek LLM 67B Base and fine-tuned on extra instruction data; DeepSeek reports around 1.5 million English and Chinese SFT instances and two SFT epochs for the 67B model. The base model was pretrained from scratch on a corpus reported as 2 trillion primarily English and Chinese tokens, and the released chat checkpoint supports a 4,096-token sequence length.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)1047
Arena Text — Chinese (No Style Control)1131
Arena Text — Creative Writing (No Style Control)1068
Arena Text — English (No Style Control)1127
Arena Text — Entertainment, Sports & Media (No Style Control)1062
Arena Text — Exclude Ties (No Style Control)943
Arena Text — Exclude Ties (Style Controlled)1055
Arena Text — Hard Prompts (No Style Control)1070
Arena Text — Hard Prompts English (No Style Control)1087
Arena Text — Instruction Following (No Style Control)1080
Arena Text — Legal & Government (No Style Control)1089
Arena Text — Life, Physical & Social Science (No Style Control)1066
Arena Text — Longer Query (No Style Control)1092
Arena Text — Medicine & Healthcare (No Style Control)1061
Arena Text — Multi-turn (No Style Control)1082
Arena Text — Non-English (No Style Control)1073
Arena Text — Non-English (Style Controlled)1149
Arena Text — Occupational Mathematical (No Style Control)1093
Arena Text — Software & IT Services (No Style Control)1096
Arena Text — Writing, Literature & Language (No Style Control)1111
Artificial Analysis Intelligence Index3.0
BBH (Open LLM Leaderboard v2)52.4
Capability110.6
GPQA (Open LLM Leaderboard v2)31.6
GPQA Diamond24.6
IFEval (Open LLM Leaderboard v2)55.9
LMArena Text Arena Chinese Elo1234
LMArena Text Arena Coding Elo1218
LMArena Text Arena Creative Writing Elo1135
LMArena Text Arena Elo1184
Loading Atlas data…