Datasets:
Multi-UNER-LLM-inst
Multi-UNER-LLM-inst is a multilingual extension of UNER-LLM-inst, the instruction-format version of Universal NER v1 (UNER) included in the Aya Collection by CohereLabs (templated_uner_llm subset).
It is an instruction-following Named Entity Recognition (NER) dataset covering English, Italian, Spanish, French and German, with aligned train/test/validation splits.
Source dataset
| Resource | Hugging Face dataset | Paper |
|---|---|---|
UNER-LLM-inst in the Aya Collection (templated_uner_llm) |
CohereLabs/aya_collection | Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning (Singh et al., 2024) |
| Universal NER instruction format | universalner/uner_llm_instructions | Universal NER: A Gold-Standard Multilingual Named Entity Recognition Benchmark (Mayhew et al., 2023) |
License of the source data: CC BY-SA 4.0
How the dataset was built
The English (17,617 samples) and German (999 samples) portions of UNER-LLM-inst were used as source data, for a total of 18,616 samples.
- English (7,448): kept in English.
- Italian, Spanish, French (2,792 each): machine-translated from distinct, non-overlapping portions of the English data.
- German (2,792): the 999 original German samples were kept as originally available. To balance the train split with the other languages, 1,793 additional English samples were machine-translated into German.
Translation models
| Language | Model | Translated samples |
|---|---|---|
| Italian | Mistral-Small-3.1-24B-Instruct-2503 |
2,792 |
| Spanish | Mistral-Small-3.1-24B-Instruct-2503 |
2,792 |
| French | Qwen3.5-122B-A10B-FP8 |
2,792 |
| German | Mistral-Small-3.1-24B-Instruct-2503 |
1,793 (the remaining 999 are original) |
Pre- and post-processing
- Entity review (pre-processing): before translation, the extracted named entities of the English source were reviewed for correctness and consistency, with manual corrections applied where needed.
- Offset adaptation (post-processing, translated samples only): after translation, the character-level
Start/Endoffsets were no longer valid. They were automatically recalculated by locating each entity's translated text in the corresponding translated sentence. - Original German samples were not modified.
- Metadata normalization (all languages):
languagewas normalized to two-letter codes,splitwas aligned with the file name,model_versionwas added.
Dataset structure
Each language is a separate configuration (en, it, es, fr, de) with train, validation and test splits.
from datasets import load_dataset
ds = load_dataset("VillanovaAI/Multi-UNER-LLM-inst", "it", split="train")
Language and split distribution
| Language | Config | Train | Validation | Test | Total | % |
|---|---|---|---|---|---|---|
| English | en |
5,151 | 933 | 1,364 | 7,448 | 40% |
| Italian | it |
1,931 | 350 | 511 | 2,792 | 15% |
| Spanish | es |
1,931 | 350 | 511 | 2,792 | 15% |
| French | fr |
1,931 | 350 | 511 | 2,792 | 15% |
| German | de |
1,931 | 350 | 511 | 2,792 | 15% |
| Total | 12,875 | 2,333 | 3,408 | 18,616 | 100% |
Fields
| Field | Description |
|---|---|
id |
Unique id of the sample as in the source dataset |
inputs |
Prompt with the task instruction, a one-shot example with its expected answer, and the final input sentence |
targets |
Expected output for the final sentence, in JSON-like format |
dataset_name |
Name of the source dataset |
sub_dataset_name |
Name of the source sub-dataset |
task_type |
Task type (named-entity-recognition) |
template_id |
Id of the template applied, as in the source dataset (1) |
language |
Two-letter language code (en, it, es, fr, de) |
split |
train, validation or test |
script |
Script of the language, as in the source dataset (Latn) |
model_version |
Model used to translate the sample. null for English samples and for the original German samples |
Entity types and output format
Entities follow the UNER scheme: PER (person), ORG (organization), LOC (location). Nationalities are annotated neither as locations nor as organizations.
targets contains a "Results" list where each entity has TypeName, Text, Start and End.
Start/End are character offsets (0-based, End exclusive) in the sentence they refer to: the example sentence for the Results block inside inputs, and the final input sentence for targets.
When the sentence has no entities, targets is "Results": [].
Data instance
{
"id": 25526,
"inputs": "Please identify all the named entities mentioned in the input sentence provided below. Use only the categories: PER - person, ORG - organization, and LOC - location.\nRemember, nationalities are neither locations nor organizations, and organizations can represent other groups of people.\nPay attention to the provided example. You should only output the results in JSON format, following a similar structure to the example result provided.\n\nExample sentence and results:\n\nWhat is this Miramar?\n\n\"Results\": [\n {\n \"TypeName\": \"LOC\",\n \"Text\": \"Miramar\",\n \"Start\": 13,\n \"End\": 20\n }\n]\n\nConsidering the input sentence below, what is the output result?\n\nDoes anyone know any good restaurants in cordoba?\n",
"targets": "\"Results\": [\n {\n \"TypeName\": \"LOC\",\n \"Text\": \"cordoba\",\n \"Start\": 41,\n \"End\": 48\n }\n]",
"dataset_name": "UNER-LLM-inst",
"sub_dataset_name": "uner-llm-inst-english",
"task_type": "named-entity-recognition",
"template_id": 1,
"language": "en",
"split": "test",
"script": "Latn",
"model_version": null
}
Limitations
- Translations are machine-generated and may contain errors or unnatural phrasing.
- Entity annotations in translated samples derive from the English annotations. Entity type ambiguity (e.g.
ORGvsLOCfor institutions, venues or cities acting as organizations) may not transfer consistently across languages. - Offsets in translated samples are computed by exact string matching; entities that were inflected or rephrased by the translation required some manual correction.
- The language distribution is deliberately unbalanced (40% English).
Licensing and attribution
This dataset is released under CC BY-SA 4.0.
It derives from UNER-LLM-inst, the instruction-format version of Universal NER v1 included in the Aya Collection. The Aya Collection as a whole is released under Apache 2.0, but its UNER-LLM-inst component is CC BY-SA 4.0. According to the Universal NER authors, this license follows from the Universal Dependencies treebanks on which the annotations were built.
Since this dataset adapts that material, it is distributed under the same license, as required by the ShareAlike clause.
Citation
If you use this dataset, please cite the Aya collection works and the original Universal NER:
@misc{singh2024aya,
title={Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning},
author={Shivalika Singh and others},
year={2024},
eprint={2402.06619},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/pdf/2402.06619}
}
@misc{mayhew2023universal,
title={Universal NER: A Gold-Standard Multilingual Named Entity Recognition Benchmark},
author={Stephen Mayhew and Terra Blevins and Shuheng Liu and Marek Šuppa and Hila Gonen and Joseph Marvin Imperial and Börje F. Karlsson and Peiqin Lin and Nikola Ljubešić and LJ Miranda and Barbara Plank and Arij Riabi and Yuval Pinter},
year={2023},
eprint={2311.09122},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/pdf/2311.09122}
}
- Downloads last month
- 2