Atlas

Models

← All models

Qwen1.5 4B Chat

Alibaba · 2024-02-04 · 4B parameters

Qwen1.5-4B-Chat is Alibaba's aligned chat checkpoint in the Qwen1.5 family, announced on February 4, 2024. It is a dense decoder-only Transformer with 3,950,369,280 parameters, a 32,768-token context window, and text-only input and output; Qwen states that the chat models were post-trained with supervised fine-tuning and direct preference optimization. Its chat template supports system, user, and assistant roles.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)965
Arena Text — Chinese (No Style Control)1023
Arena Text — Creative Writing (No Style Control)970
Arena Text — English (No Style Control)1015
Arena Text — Entertainment, Sports & Media (No Style Control)953
Arena Text — Exclude Ties (No Style Control)764
Arena Text — Exclude Ties (Style Controlled)895
Arena Text — Expert (No Style Control)979
Arena Text — German (No Style Control)902
Arena Text — Hard Prompts (No Style Control)976
Arena Text — Hard Prompts English (No Style Control)985
Arena Text — Instruction Following (No Style Control)979
Arena Text — Legal & Government (No Style Control)1010
Arena Text — Life, Physical & Social Science (No Style Control)977
Arena Text — Longer Query (No Style Control)988
Arena Text — Medicine & Healthcare (No Style Control)955
Arena Text — Multi-turn (No Style Control)978
Arena Text — Non-English (No Style Control)979
Arena Text — Non-English (Style Controlled)1073
Arena Text — Occupational Mathematical (No Style Control)998
Arena Text — Russian (No Style Control)951
Arena Text — Software & IT Services (No Style Control)996
Arena Text — Writing, Literature & Language (No Style Control)1001
BBH (Open LLM Leaderboard v2)40.1
Capability86.4
Creative Writing v227.0
EQ-Bench v228.8
EQ-Bench v2 + MAGI-Hard Combined30.7
GPQA (Open LLM Leaderboard v2)26.7
IFEval (Open LLM Leaderboard v2)31.6
Loading Atlas data…