This repository is the official implementation of Benchmarking Complex Instruction-Following with Multiple Constraints Composition. Our Paper
To install requirements:
pip install -r requirements.txt
All data of ComplexBench is placed in data/data_release.json . The entire dataset is organized in the format of a list, where each element of the list is an instance of the dataset. The format of an instance is as follows.
-
main_id(integer): A unique identifier for the instance. -
group(string): The task group identifier for The Coherent Test for Selection (Section 5.2.3). Applicable only in instructions that contain Selection. -
idx_in_group(integer): Number of the instruction in the task group. Applicable only in instructions that contain Selection. -
instruction(string): The actual instruction. -
task_types(string): The task type of the instruction. -
constraint_dimensions(list): All constraint dimensions within the instruction. -
composition_types(list): All composition types of constraints within the instruction. -
category(list): The category of the instruction based on composition types within it. -
scoring_questions(list): All scoring questions of the instruction, which are designed to verify each constraint dimensions and composition types. Each element of this list contains:point_id(integer): Number of the scoring question.question(string): The actual scoring question.rule(string): Verification rule for the score scoring.nullindicates that the score question cannot be verified by rules.constraint_dimensions(string): All constraint dimensions that the scoring question verifies.composition_types(string): All composition types that the scoring question verifies.dep(list): The numbers of all score questions on which this score question depends..
-
sub_instructions(dict): The decomposed atomic instrcutions based on the instruction (Section 5.2.3). All keys in this dict are in the format ofsub_instruction_x, wherexis the number of the decomposed instruction. Each value of this dict contains:instruction(string): The actual decomposed atomic instruction.scoring_questions(list): All scoring questions of the decomposed atomic instrcution .
Here is an example of ComplexBench.
{
"main_id": 899,
"group": "complex_instruction_eval_1285",
"idx_in_group": 1,
"instruction": "依次判断以下两个案例中的国家是否有特别提款权。如果有，请写出一篇为该国申请提款的文章，字数不少于300字，且分点明确。如果没有则解释原因，字数不超过100字。\n\n案例1：\n\n国家A是一个发展中国家，是国际货币基金组织（IMF）的成员国，正在经历一场自然灾害，导致该国经济陷入危机，失去了国际支付能力。\n\n案例2：\n\n国家B是一个富裕国家，面临着国内通货膨胀问题，B是IMF的成员国，拥有充足的外汇储备。",
"task_types": "Professional Writing",
"constraint_dimensions": [
"Length",
"Helpfulness",
"Bullets Format",
"Factuality"
],
"composition_types": [
"And",
"Selection"
],
"category": "Selection_2",
"scoring_questions": [
{
"point_id": 0,
"question": "模型是否正确判断国家A有特别提款权？",
"rule": null,
"constraint_dimensions": [
"Factuality"
],
"composition_types": [
"Selection"
],
"dep": []
},
{
"point_id": 1,
"question": "模型是否正确判断国家B没有特别提款权？",
"rule": null,
"constraint_dimensions": [
"Factuality"
],
"composition_types": [
"Selection"
],
"dep": []
},
{
"point_id": 2,
"question": "模型是否根据国家A有特别提款权生成申请提款的文章？",
"rule": null,
"constraint_dimensions": [
"Helpfulness"
],
"composition_types": [],
"dep": [
0
]
},
{
"point_id": 3,
"question": "模型生成的申请提款文章是否逻辑合理，符合事实？",
"rule": null,
"constraint_dimensions": [
"Factuality"
],
"composition_types": [],
"dep": [
0
]
},
{
"point_id": 4,
"question": "模型生成的申请提款文章是否在300字以上？",
"rule": "model_length:[300,10000]",
"constraint_dimensions": [
"Length"
],
"composition_types": [],
"dep": [
0
]
},
{
"point_id": 5,
"question": "模型生成的申请提款文章是否分点明确？",
"rule": null,
"constraint_dimensions": [
"Bullets Format"
],
"composition_types": [],
"dep": [
0
]
},
{
"point_id": 6,
"question": "模型是否生成国家B没有特别提款权的解释？",
"rule": null,
"constraint_dimensions": [
"Helpfulness"
],
"composition_types": [],
"dep": [
1
]
},
{
"point_id": 7,
"question": "模型生成的解释是否逻辑合理，符合事实？",
"rule": null,
"constraint_dimensions": [
"Factuality"
],
"composition_types": [],
"dep": [
1
]
},
{
"point_id": 8,
"question": "模型生成的解释是否不超过100字？",
"rule": "model_length:[1,100]",
"constraint_dimensions": [
"Length"
],
"composition_types": [],
"dep": [
1
]
}
],
"sub_instructions": {
"sub_instruction_0": {
"instruction": "判断以下两个案例中的国家是否有特别提款权。\n\n案例1：\n\n国家A是一个发展中国家，是国际货币基金组织（IMF）的成员国，正在经历一场自然灾害，导致该国经济陷入危机，失去了国际支付能力。\n\n案例2：\n\n国家B是一个富裕国家，面临着国内通货膨胀问题，B是IMF的成员国，拥有充足的外汇储备。",
"scoring_questions": [
"模型是否正确判断国家A有特别提款权？",
"模型是否正确判断国家B没有特别提款权？"
]
},
"sub_instruction_1": {
"instruction": "请根据上面的判断，完成下面的指令。\n\n- 如果有，请写出一篇为该国申请提款的文章，字数不少于300字，且分点明确。\n- 如果没有则解释原因，字数不超过100字。",
"scoring_questions": [
"模型是否根据国家A有特别提款权生成申请提款的文章？",
"模型生成的申请提款文章是否逻辑合理，符合事实？",
"模型生成的申请提款文章是否在300字以上？",
"模型生成的申请提款文章是否分点明确？",
"模型是否生成国家B没有特别提款权的解释？",
"模型生成的解释是否逻辑合理，符合事实？",
"模型生成的解释是否不超过100字？"
]
}
}
}First, you need to deploy your target LLM and generate responses of it (This part is not included in this repository). We have placed the generated responses of 15 LLMs in our paper in llm_generations. You can refer the format of one of these files and replace genereated field with your target LLM responses.
We suggest using greedy decoding to avoid the randomness of decoding.
Then, you can evaluate any desired model via scirpt eval.sh:
bash eval.sh "<path_to_data_release.json>" "<path_to_generated_responsed_in_step_1>" "<path_to_results>" "<openai_API_KEY>" "<openai_API_KEY>" "<your_model_name>"Here is an example:
bash eval.sh "data/data_release.json" "llm_generations/glm4.jsonl" "evaluation_results" "<openai_API_KEY>" "<openai_API_KEY>" "glm4"In this example, you can find the results of each scored question in evaluation_results/glm4_final_results.json. The statistical data of the final results will be saved in evaluation_results/glm4_statistics.json.
@article{wen2024benchmarking,
title={Benchmarking Complex Instruction-Following with Multiple Constraints Composition},
author={Wen, Bosi and Ke, Pei and Gu, Xiaotao and Wu, Lindong and Huang, Hao and Zhou, Jinfeng and Li, Wenchuang and Hu, Binxin and Gao, Wendy and Xu, Jiaxin and others},
journal={arXiv preprint arXiv:2407.03978},
year={2024}
}
Please kindly cite our paper if this paper and the codes are helpful.