VillanovaAI/Multi-Persona-IF

Multilingual Persona IF Dataset

Overview

Multilingual Persona IF is a multilingual version of the English dataset Persona IF, create by AllenAI as part of the Tulu3 dataset suite. The original Persona IF dataset is also a synthetic dataset, created for instruction-tuning of Large Language Models. It has a particular focus on satisfying constraints imposed by the user in the request to the assistant, and as such borrows a list of constraints from the benchmarking dataset IFEval.

The pipeline for generating the dataset is described in the AllenAI paper about their open model Tulu3. Their code can be found in this repository. They started from a list of constraints, then manually wrote some seed instructions for each of the constraints. They then prompted a LLM to generate an instruction containing 1-3 randomly chosen constraints, using the seeds as few-shot examples. Once the instructions were generated, they asked the same LLM to also provide the assistant's responses to the instructions.

In order to create some variety in the instructions, they used the Persona Hub personas by Ge et al., 2024. When prompting the LLM, they asked to generate an instruction that the given persona might ask an assistant to do.

To generate the multilingual version of Persona IF, we employed the same methodology and pipeline as AllenAI, but we used the permissively licensed LLM Qwen3.5-122B-A10B and VillanovaAI's multilingual Persona dataset. We also reworked and translated the list of constraints, seed instructions and prompts to fit our needs.

Language Distribution

The original dataset contains 29,980 English instructions. Since we wouldn't be translating the original dataset, we chose to generate 15,000 instructions for each of our five languages for a total of 75,000 instructions. The final distribution is as follows:

Language	Subset Name	Percentage	Number of Examples
English	en	20%	15,000
Italian	it	20%	15,000
French	fr	20%	15,000
Spanish	es	20%	15,000
German	de	20%	15,000

Additional information

When we examined the original list of constraints, we noticed that some constraints were logically incompatible with each other and would result in an instruction that would be effectively impossible to follow. An example of this situation are the two constraint types "in lowercase" and "number of capital words": it is impossible to write a text in lowercase that containts a number of capital words other than 0. Therefore we modified the randomized constraints sampling pipeline to avoid such conflicts.

Dataset Structure

Each entry contains the following fields:

Field	Type	Description
id	string	Unique index of the instruction
instruction	string	Text of the generated instruction
input_persona	string	The persona used to prompt the LLM when generating the instruction
response	string	Text of the assistant's response
given_constraints	list	A list of 1 to 3 randomly chosen constraints to be followed

Licensing

The dataset is distributed under the ODC-BY-1.0 license. Users must adhere to the attribution and usage requirements outlined under this license.