datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
gliner-opf-ptbr-pii-bench-v1
PT-BR PII Benchmark v1 — head-to-head
Why this exists
Open-sourced at the request of @arthrod (Arthur Souza Rodrigues) after a two-night sprint training and benchmarking these models on an AMD MI300X. The motivation: there's surprisingly little published head-to-head data comparing MoE-based PII detectors (openai/privacy-filter) against dense small-model approaches (GLiNER on mmBERT/ettin) on a real-world Portuguese task — and the trade-offs turned out to be sharp enough to be… See the full description on the dataset page: https://huggingface.co/datasets/arthrod/gliner-opf-ptbr-pii-bench-v1.gliner_review_comparisongliner-biomed-pre-training
GLiNER-BioMed pre-training dataset
This dataset, used for the pre-training stage of the GLiNER-BioMed models, was introduced in the paper GLiNER-BioMed: a suite of efficient models for open biomedical named entity recognition.
Citation
If you use the GLiNER-BioMed models or datasets in your work, please cite:
@article{yazdani2026gliner,
author = {Yazdani, Anthony and Stepanov, Ihor and Teodoro, Douglas},
title = {{GLiNER-BioMed}: a suite of efficient… See the full description on the dataset page: https://huggingface.co/datasets/anthonyyazdaniml/gliner-biomed-pre-training.gliner2-fr-biomed-v3e-trainset
gliner2-fr-biomed-v3e-trainset
Données d'entraînement synthétiques (distillation Qwen3-235B) du modèle rntc/gliner2-fr-biomed-v3e — GLiNER2 FR biomédical (primitives génériques + role-binding). Aucun gold de benchmark (préserve la généralisation zero-shot).
Format JSONL, une ligne = un exemple :
{"input": "<texte>", "output": {"entities": {"<label>": ["<mention>", ...]},
"classifications": {...}, "json_structures": [{"<section>": {"<champ>": "<valeur>"}}], "relations": [...]}}… See the full description on the dataset page: https://huggingface.co/datasets/rntc/gliner2-fr-biomed-v3e-trainset.gliner-biomed-post-training
GLiNER-BioMed post-training dataset
This dataset, used for the post-training stage of the GLiNER-BioMed models, was introduced in the paper GLiNER-BioMed: a suite of efficient models for open biomedical named entity recognition.
Citation
If you use the GLiNER-BioMed models or datasets in your work, please cite:
@article{yazdani2026gliner,
author = {Yazdani, Anthony and Stepanov, Ihor and Teodoro, Douglas},
title = {{GLiNER-BioMed}: a suite of efficient… See the full description on the dataset page: https://huggingface.co/datasets/anthonyyazdaniml/gliner-biomed-post-training.gliner-biomed-curated-corpus
GLiNER-BioMed curated corpus
Unlabeled corpus introduced in the paper GLiNER-BioMed: A Suite of Efficient Models for Open Biomedical Named Entity Recognition.
Citation
If you use the GLiNER-BioMed models or datasets in your work, please cite:
@misc{yazdani2025glinerbiomedsuiteefficientmodels,
title={GLiNER-BioMed: A Suite of Efficient Models for Open Biomedical Named Entity Recognition},
author={Anthony Yazdani and Ihor Stepanov and Douglas Teodoro}… See the full description on the dataset page: https://huggingface.co/datasets/anthonyyazdaniml/gliner-biomed-curated-corpus.tool-output-extraction-swebench-gliner
Tool Output Extraction (extractive / GLiNER2 format)
Extractive variant of KRLabsOrg/tool-output-extraction-swebench, formatted for fine-tuning span-extraction models (GLiNER2, BERT-for-QA, etc.).
Each tool observation from the parent dataset is chunked into ~400-token windows (preserving line boundaries) so it fits into encoder-style models with a 512-token context. The query is concatenated in front of each chunk, extractive-QA style, and gold evidence is mapped to verbatim spans… See the full description on the dataset page: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench-gliner.tool-output-extraction-swebench-gliner2-v2
Tool Output Extraction — GLiNER2 line-number format (experimental)
Experimental variant of KRLabsOrg/tool-output-extraction-swebench, reformulated for GLiNER2's architectural constraints.
GLiNER2 has max_width=8 (max entity length = 8 tokens), which is incompatible with the multi-line evidence blocks in the parent dataset (50–500 tokens). This variant extracts line-number prefixes ("233:") as short entities instead, keeping the full chunk as context.
For general extractive training… See the full description on the dataset page: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench-gliner2-v2.gliner2-fr-biomed-v3b-trainset
gliner2-fr-biomed-v3b-trainset
Données d'entraînement synthétiques (distillation Qwen3-235B) du modèle rntc/gliner2-fr-biomed-v3b — GLiNER2 FR biomédical (généraliste (UMLS-steered, multi-task)). Aucun gold de benchmark (préserve la généralisation zero-shot).
Format JSONL, une ligne = un exemple :
{"input": "<texte>", "output": {"entities": {"<label>": ["<mention>", ...]},
"classifications": {...}, "json_structures": [{"<section>": {"<champ>": "<valeur>"}}], "relations":… See the full description on the dataset page: https://huggingface.co/datasets/rntc/gliner2-fr-biomed-v3b-trainset.gliner2-fr-biomed-v3f-trainset
gliner2-fr-biomed-v3f-trainset
Données d'entraînement synthétiques (distillation Qwen3-235B) du modèle rntc/gliner2-fr-biomed-v3f — GLiNER2 FR biomédical (densité stratifiée par hash (100k chunks annotés)). Aucun gold de benchmark (préserve la généralisation zero-shot).
Format JSONL, une ligne = un exemple :
{"input": "<texte>", "output": {"entities": {"<label>": ["<mention>", ...]},
"classifications": {...}, "json_structures": [{"<section>": {"<champ>": "<valeur>"}}]… See the full description on the dataset page: https://huggingface.co/datasets/rntc/gliner2-fr-biomed-v3f-trainset.gliner-multilingual-ner-silver-v1
Betterdata Annotated Multilingual NER/PII
Dataset Summary
This dataset contains multilingual, annotated NER/PII spans across 13 languages with 60+ label classes spanning PII, PHI, PCI, and general entity types. It is designed to train and evaluate privacy-preserving NER models.
Data Sources
bloomberg_financial_news_annotated (data/augmented/bloomberg_financial_news_annotated.jsonl)
c4_multilingual_annotated (data/augmented/c4_multilingual_annotated.jsonl)… See the full description on the dataset page: https://huggingface.co/datasets/betterdataai/gliner-multilingual-ner-silver-v1.gliner2-ptbr-ontoevidence-data
OntoEvidence-BR
OntoEvidence-BR is an open Brazilian Portuguese dataset for GLiNER, GLiNER2, NER, schema-guided information extraction, ontology-guided extraction, and operational service triage.
OntoEvidence-BR é um dataset aberto em português brasileiro para extração de evidências operacionais orientadas por ontologia, com frases curtas, ruidosas e hard negatives semânticos.
Descrição
Tamanho: ~2.014 amostras (train: 1.812 / val: 100 / test: 102)
Licença:… See the full description on the dataset page: https://huggingface.co/datasets/ottema/gliner2-ptbr-ontoevidence-data.gliner2-fr-biomed-v3c-trainset
gliner2-fr-biomed-v3c-trainset
Données d'entraînement synthétiques (distillation Qwen3-235B) du modèle rntc/gliner2-fr-biomed-v3c — GLiNER2 FR biomédical (densité bimodale + relations). Aucun gold de benchmark (préserve la généralisation zero-shot).
Format JSONL, une ligne = un exemple :
{"input": "<texte>", "output": {"entities": {"<label>": ["<mention>", ...]},
"classifications": {...}, "json_structures": [{"<section>": {"<champ>": "<valeur>"}}], "relations": [...]}}… See the full description on the dataset page: https://huggingface.co/datasets/rntc/gliner2-fr-biomed-v3c-trainset.gliner-biomed-balanced-curated-corpus
GLiNER-BioMed balanced curated corpus
Balanced, unlabeled corpus introduced in the paper GLiNER-BioMed: A Suite of Efficient Models for Open Biomedical Named Entity Recognition.
Citation
If you use the GLiNER-BioMed models or datasets in your work, please cite:
@misc{yazdani2025glinerbiomedsuiteefficientmodels,
title={GLiNER-BioMed: A Suite of Efficient Models for Open Biomedical Named Entity Recognition},
author={Anthony Yazdani and Ihor Stepanov and Douglas… See the full description on the dataset page: https://huggingface.co/datasets/anthonyyazdaniml/gliner-biomed-balanced-curated-corpus.GLINER-multi-task-synthetic-dataThis is official synthetic dataset used to train GLiNER multi-task model.
The dataset is a list of dictionaries consisting a tokenized text with named entity recognition (NER) information. Each item represents of two main components:
'tokenized_text': A list of individual words and punctuation marks from the original text, split into tokens.
'ner': A list of lists containing named entity recognition information. Each inner list has three elements:
Start index of the named entity in the… See the full description on the dataset page: https://huggingface.co/datasets/knowledgator/GLINER-multi-task-synthetic-data.GLiNER-FUNSDDataset derived from FUNSD for training GLiNER-based multi-modal models.
Processing script:
from datasets import load_dataset
import shutil
import json
import os
dataset = load_dataset("nielsr/funsd")
raw_labels = dataset['train'].features['ner_tags'].feature.names
gliner_pdf_dataset = []
output_dir = "gliner_funsd"
os.makedirs(output_dir, exist_ok=True)
os.makedirs(os.path.join(output_dir, 'images'), exist_ok=True)
def process_dataset(example):
tokens = example['words']
bboxes =… See the full description on the dataset page: https://huggingface.co/datasets/Ihor/GLiNER-FUNSD.eval-gliner2-ner-fin-dice-soft-20260709eval-gliner2-fold3-20260708eval-gliner2-ner-classweight-3seed
Class-weighted loss — 3 seed x 4 dataset
train: --loss-class-weight tren fastino/gliner2-multi-v1, batch 16, patience 2, seeds [42, 43, 44]
eval : --threshold 0.7 --use-desc --schema-case original --no-extra-val (giong loss ablation)
dataset
epochs
eval split
strict macro F1
lenient micro F1
conll2003
6
test
86.06 +/- 0.95
89.31 +/- 0.95
ontonotes5
3
test
77.47 +/- 0.51
90.30 +/- 0.06
fin
8
test
38.15 +/- 2.40
75.92 +/- 1.36
mit_restaurant
6
test
81.55 +/-… See the full description on the dataset page: https://huggingface.co/datasets/AITeamUIT/eval-gliner2-ner-classweight-3seed.eval-gliner2-ner-bionlp2004-boundary-smoothing-validationeval-gliner2-ner-setconsistent
Set-consistent — 4 bo NER, 18 run
base model: fastino/gliner2-multi-v1
train: --loss set (mac dinh cua train.py cho moi tham so set)
eval: --threshold 0.5 --use-desc --schema-case original --no-extra-val
--loss set tu bat learnable null anchor trong scorer, nen checkpoint decode o
threshold 0.5 — khac 0.7 cua bang loss-ablation 5 config. Khac biet co chu
dich, khong phai threshold tuning.
Ngan sach tung bo
bo
repo
nhan
split
seeds
epochs
batch
fin… See the full description on the dataset page: https://huggingface.co/datasets/AITeamUIT/eval-gliner2-ner-setconsistent.gliner-funding-preparedeval-gliner2-mdeberta_base-uni-20260621eval-gliner2-fold2-20260708gliner-bird-diet-synthetic
GLiNER Bird Diet Synthetic Dataset
This is an NER dataset focused on ornithological data, specifically focused on the diets of birds. The data is purely synthetic and should not be taken as factual. We created this dataset using Qwen2-7B-Instruct. It consists of ~2k descriptions. The format of the annotations consists with the GLiNER format. We used this data to finetune a GLiNER model. For the base model, we used NuNerZero Span. You can visit our model here: GLiNER Ecology Diet.… See the full description on the dataset page: https://huggingface.co/datasets/wjbmattingly/gliner-bird-diet-synthetic.eval-gliner2-modernbert_pasteproof-uni-20260621eval-gliner2-fold0-20260708eval-gliner2-ner-fin-lossablation
fin — loss ablation (5 config x 3 seed)
base model: fastino/gliner2-multi-v1
dataset: quynong/fin (4 nhan), eval tren split test
train: 8 epochs, batch 16, early-stopping patience 2, seeds [42, 43, 44]
eval: --threshold 0.7 --use-desc --schema-case original --no-extra-val (giong nhau cho ca 5 config)
Ket qua (micro F1 %, mean +/- std tren seed)
mode
config
n seeds
micro F1
macro F1
micro P
micro R
lenient
bce
3
77.88 +/- 4.74
41.96 +/- 8.81
88.63… See the full description on the dataset page: https://huggingface.co/datasets/AITeamUIT/eval-gliner2-ner-fin-lossablation.eval-gliner2-ner-ontonotes5-boundary-smoothing-testeval-gliner2-ner-fin-boundary-smoothing-validation