Overview

评估大型语言模型严格遵循指令的能力，包含500个可验证的指令

Metrics

Metric	Unit	Direction
prompt_strict_acc	%	↑ Higher is better
prompt_loose_acc	%	↑ Higher is better

Sources

Model Score Ranking

#	Model	Vendor	Score
1	Gemini 1.5 Pro 002	google	87.9
2	GPT-4o (2024-05-13)	openai	86.7
3	Gemini 2.0 Flash	google	85.9
4	GPT-4o (2024-08-06)	openai	85.5
5	o1 Preview	openai	84.4
6	Qwen1.5 110B	alibaba	83.4
7	Sonar Reasoning	other	83.1
8	Gemini 2.0 Flash Thinking	google	82.8
9	o1	openai	82.2
10	Claude 3.5 Sonnet (2024-10-22)	anthropic	81.6
11	Gemini 1.0 Ultra	google	81.6
12	Yi Vision	other	81.4
13	Command Nightly	cohere	80.8
14	StableLM 2 12B	other	80
15	Qwen1.5 14B	alibaba	79.9
16	DBRX Base	other	79.3
17	Llama 3.3 70B	meta	79.3
18	Yi 1.5 34B	other	79.3
19	Claude 3 Opus	anthropic	79.2
20	GLM-4 Plus	other	79.2
21	Jamba Instruct	other	79.2
22	Sonar Huge	other	78.7
23	DeepSeek V2	deepseek	78.6
24	Hermes 3 Llama 3.1 405B	other	78.2
25	GPT-4 Turbo	openai	77.9
26	Mistral Medium	mistral	77.4
27	Gemini 1.5 Flash 002	google	77.2
28	Grok-2 Vision	xai	76.9
29	Mistral Small 3	mistral	76.9
30	Qwen2 72B	alibaba	76.9
31	DeepSeek LLM 67B	deepseek	76.8
32	Qwen1.5 72B	alibaba	76.4
33	Mistral Large	mistral	75.8
34	GLM-4 Air	other	75.7
35	Claude 3 Opus (2024-02-29)	anthropic	75.5
36	Claude 3 Sonnet (2024-02-29)	anthropic	75.5
37	Claude 3 Haiku (2024-03-07)	anthropic	75.3
38	Microsoft WizardLM 2 8x22B	other	75.3
39	Nous Hermes 2 Mixtral 8x7B	other	75.3
40	Gemini 1.0 Pro	google	75.1
41	Sonar Large	other	74.9
42	GPT-4 32K	openai	74.7
43	GPT-4	openai	74.6
44	Llama 3.1 Nemotron 70B	meta	74.5
45	Llama 3.2 90B Vision	meta	74.5
46	Claude 3 Sonnet	anthropic	74.3
47	Gemini 1.5 Flash-8B 002	google	74.3
48	Phi-3.5 MoE	other	74.3
49	Orca 2 13B	other	74.2
50	Command R	cohere	74.1
51	Qwen2.5 14B	alibaba	74.1
52	Mistral Small	mistral	74
53	GPT-4 1106 Preview	openai	73.9
54	Hermes 3 Llama 3.1 70B	other	73.8
55	Llama 3.1 70B	meta	73.7
56	Claude 3.5 Haiku	anthropic	73.6
57	GPT-4 Vision	openai	73.5
58	Grok-2 Mini	xai	73.5
59	Qwen2.5 32B	alibaba	73.5
60	Yi Large	other	73.4
61	DBRX Instruct	other	73.1
62	o1 mini	openai	73.1
63	GPT-4 Vision Preview	openai	72.7
64	Gemma 2 27B	google	72.5
65	Llama 3 70B	meta	72.2
66	Command R+ (08-2024)	cohere	72.1
67	Mixtral 8x7B	mistral	72.1
68	Sonar Small	other	72.1
69	Jamba 1.5 Mini	other	71.9
70	Gemini 1.0 Flash	google	71.8
71	DeepSeek V2 Chat	deepseek	71.5
72	GPT-4 0125 Preview	openai	71.4
73	Qwen2 57B	alibaba	71.2
74	Claude 3 Haiku	anthropic	70.8
75	Jamba 1.5	other	70.8
76	Jamba 1.5 Large	other	70.8
77	WizardLM Team WizardLM 2 8x22B	other	70.7
78	Yi Large Turbo	other	70.7
79	Falcon 180B	other	70.5
80	Command R (08-2024)	cohere	70.3
81	Zephyr ORPO 141B Alpha	other	69.7
82	NVIDIA Llama 3.1 Nemotron 70B	other	69.6
83	OLMo 7B	other	68.4
84	Nous Hermes 2 Yi 34B	other	68.3
85	Llama 3 8B	meta	68.2
86	Qwen1.5 32B	alibaba	68.2
87	Gemini 1.5 Flash-8B	google	68.1
88	Gemini 1.5 Flash	google	68
89	GPT-4o mini	openai	67.9
90	GLM-4V 9B	other	67.4
91	Phi-3 Medium	other	67.4
92	Command R7B	cohere	67.2
93	Gemma 7B	google	66.9
94	GLM-4 Flash	other	65.9
95	Codestral Mamba	mistral	65.5
96	Yi 1.5 9B	other	65.4
97	Gemma 2 9B	google	64.3
98	Code Llama 70B	meta	63.8
99	Mathstral 7B	mistral	63.6
100	OLMo 7B Instruct	other	63.6
101	StarChat2 15B v0.1	other	63.3
102	Llama 3.1 8B	meta	62.1
103	Qwen2 7B	alibaba	61.8
104	Code Llama 7B	meta	61.2
105	OLMo 1.7 7B	other	61.2
106	DeepSeek Math 7B	deepseek	61.1
107	Phi-3 Small	other	61
108	Mistral 7B v0.3	mistral	60.9
109	Hermes 3 Llama 3.1 8B	other	60.7
110	GLM-4 9B Chat	other	59.9
111	MPT 7B	other	59.9
112	Phi-4	other	59.6
113	StarCoder2 15B	other	59.6
114	Text Bison	google	59.5
115	Codestral	mistral	59.4
116	Code Bison	google	59.2
117	Mistral 7B v0.2	mistral	59.2
118	Mistral 7B v0.1	mistral	58.9
119	DeepSeek Coder 33B	deepseek	58.8
120	Yi 1.5 6B	other	58.8
121	Code Llama 34B	meta	58.7
122	Llama 2 7B	meta	58.5
123	Llama 3.2 11B Vision	meta	58.4
124	Microsoft WizardMath 7B v1	other	57.7
125	GPT-3.5 Turbo	openai	57.4
126	Qwen2 1.5B	alibaba	57.4
127	Falcon 40B	other	57.2
128	Llama 2 70B	meta	57.1
129	Microsoft WizardLM 2 7B	other	56.9
130	DeepSeek Coder V2	deepseek	56.7
131	OpenChat 3.6 8B	other	56.7
132	Phi-3 Vision	other	56.4
133	Grok Beta	xai	56.2
134	OLMo 7B SFT	other	56.2
135	Ministral 8B	mistral	56.1
136	Mistral Nemo	mistral	56
137	OLMo 2 1124 7B	other	56
138	Qwen2.5 3B	alibaba	56
139	Flan-T5 XL	other	55.7
140	TigerBot 70B Chat	other	55.7
141	Flan-UL2	other	55.6
142	Nous Hermes 2 Solar 10.7B	other	55.6
143	DeepSeek Coder 7B	deepseek	55.4
144	Qwen2.5 7B	alibaba	55.4
145	StableLM Zephyr 3B	other	55.3
146	StarCoder2 3B	other	55.2
147	Claude 2	anthropic	55
148	Claude 2.1	anthropic	55
149	Phi-3 Mini	other	54.9
150	Yi 6B	other	54.9
151	StarCoder2 7B	other	54
152	Command Light	cohere	53.9
153	Microsoft WizardCoder Python 34B	other	53.8
154	Phi-1	other	53.8
155	Claude Instant 1	anthropic	53.2
156	Llemma 7B	other	52.9
157	Nous Capybara 34B	other	52.9
158	GPT-3.5	openai	52.8
159	Capybara 1.5B	other	52.4
160	StableCode 3B	other	51.8
161	Mistral Tiny	mistral	51.7
162	Llama Guard 2 8B	meta	51.6
163	StableLM 3 4B	other	51.6
164	Flan-T5 XXL	other	51.3
165	MPT 30B	other	51.2
166	Zephyr 7B Beta	other	51.2
167	Zephyr 7B Alpha	other	50.9
168	OpenChat 3.5 1210	other	50.4
169	Baichuan2 7B Chat	other	50.1
170	Code Llama 13B	meta	50.1
171	Llama Guard 3 8B	meta	50
172	Llama 2 13B	meta	49.9
173	Jurassic-2 Mid	other	49.8
174	Phi-2	other	48.8
175	Phi-3.5 Mini	other	48.7
176	Llama 3.2 3B	meta	48.6
177	Baichuan2 13B Chat	other	48.5
178	GPT-3.5 Turbo 16K	openai	48.5
179	Qwen2.5 0.5B	alibaba	48.5
180	Chat Bison	google	48.2
181	Yi 34B	other	48.2
182	Phi-1.5	other	48
183	Jurassic-2 Ultra	other	47.9
184	Ministral 3B	mistral	47.8
185	Grok Vision Beta	xai	46.8
186	Gemma 2B	google	46.4
187	Open-Platypus	other	46
188	PaLM 2	google	44.7
189	Nous Capybara 7B	other	44.2
190	Llama 3.2 1B	meta	44.1
191	StableLM 2 1.6B	other	44.1
192	Argilla Notus 7B v1	other	42.8
193	ChatGLM3 6B	other	42.5
194	Qwen2.5 1.5B	alibaba	41.9
195	Embed English v3	cohere	0

IFEval

설명

评估大型语言模型严格遵循指令的能力，包含500个可验证的指令

핵심 사양

카테고리	라이선스	최종 업데이트
reasoning	Apache-2.0	2024-02-01

벤치마크

단위
%
%

소스

공식 URL

https://github.com/google-research/google-research/tree/master/instruction_following_eval