Verifiable Format Control for Large Language Model Generations

The VFF dataset contains 64 unique meta constraints for LLM format control generations, where these constraints cover common needs for LLM generation output formats (e.g., JSON output). Please find the details in the paper "Verifiable Format Control for Large Language Model Generations", Findings of NAACL 2025.

Dataset Structure

Data Instances

For each instance, there is an instruction string, an input string (optional), a list of decomposed questions, and a list of the labels for each decomposed question.

{"uuid": 4, "category": "Limited grammer", "constraint": "The [[VAR1]] sentence must start with a [[VAR2]].", "vars": [{"name": "VAR1", "type": "int", "values": [1, 2, 3, 4, 5]}, {"name": "VAR2", "type": "string", "values": ["noun", "adjective", "verb", "adverb"]}], "verify": "Check if the specified sentence starts with the required part of speech."}

Data Fields

uuid: a int number representing the meta constraint index.
category: a string containing containing Limited Word Count or Limited Content or Specific Number Format , Limited Grammar or Limited Structure or Limited Punctuation.
constraint: a string containing description of this meta constraint.
vars: a list with constraint variables,each variable is a dictionary containing the variable's name, type, possible values, and a description of how to verify it.

Dataset Usage

You can use datasets such as alpaca to build the data you need.

Instruction: Begin a story with a protagonist who is the CEO of an international corporation.

constraint: Include at least [[VAR1]] of the following keywords in your response: [[VAR2]].

uuid:35

[[VAR1]]: [1, 2, 3, 4, 5]

[[VAR2]]: [['Eiffel Tower', 'Louvre Museum', 'Notre Dame', 'Seine River', 'Montmartre'], ['breakfast', 'lunch', 'dinner', 'accommodation', 'transportation'], ['museum', 'park', 'historical site', 'shopping district', 'local cuisine']]

Variables are selected from the variable tables of [[VAR1]] and [[VAR2]]

Instruction with constraint: Begin a story with a protagonist who is the CEO of an international corporation. Please observe the following constraint--Include at least 2 of the following keywords in your response: ['Eiffel Tower', 'Louvre Museum', 'Notre Dame', 'Seine River', 'Montmartre'].

After the LLM gives a response, call verify, and the function corresponding to the uuid in py verifies it.The parameters passed are: LLM-response-text,args,uuid

args for example:[2,['Eiffel Tower', 'Louvre Museum', 'Notre Dame', 'Seine River', 'Montmartre']]

Citation Information

If you find our data or codes useful, please kindly cite

@inproceedings{wang-etal-2025-verifiable,
    title = "Verifiable Format Control for Large Language Model Generations",
    author = "Wang, Zhaoyang  and
      Jiang, Jinqi  and
      Zhou, Huichi  and
      Zheng, Wenhao  and
      Zhang, Xuchao  and
      Bansal, Chetan  and
      Yao, Huaxiu",
    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
    year = "2025"
}

Datasets:

jinqij
/

VFF