Multilingual MentorES is a multilingual version of the Spanish-language dataset MentorES. The original MentorES dataset is a manually curated, open-source dataset for instruction-tuning of Large Language Models. The following tasks are represented in the dataset: closed QA, open QA, general QA, classification, information extraction, summarization, creative writing and brainstorming.

The multilingual version was obtained by translating parts of the original dataset respectively into French, German and Italian. The models used for translation were Qwen3-VL-235B-A22B-Instruct for French and Italian, Mistral-Small-3.1-24B-Instruct for German. Both models have permissive licenses that allow reuse of their output for a variety of purposes.

Language Distribution

The original dataset contains 10,175 instructions. They were divided equally among the four represented languages, making sure to also balance the number of examples for each task across languages. The final distribution is as follows:

Language	Subset Name	Percentage	Number of Examples
Italian	it	25%	2,548
French	fr	25%	2,534
Spanish	es	25%	2,552
German	de	25%	2,541

Further processing

During the translation phase, language experts noticed localization issues in the models' output. The original dataset is understandably focused on Spanish culture and language, and the translation models sometimes could not correctly transpose the examples into the target language. This was often the case with instructions regarding aspects of the Spanish language (e.g. "List all words containing ñ in the following paragraph"), or references to Spanish titles of books, movies, plays, etc.

We isolated these problematic entries using keywords and proceeded to do a manual review of localization across the translated portions of the dataset. Overall the review involved about 450 instructions each for German, French and Italian.

Dataset Structure

Each entry contains the following fields:

Field	Type	Description
`id`	string	Unique index of the instruction
`category`	string	Type of task (QA, summarization, writing, etc)
`instruction`	string	Text of the instruction
`context`	string	Context provided to the assistant as support information
`response`	string	Text of the response
`lang`	string	Language of the entry

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.

Downloads last month: 4

Total file size:

3.82 MB