SauerkrautLM-GLiNER

SauerkrautLM-GLiNER is a multilingual GLiNER-style model for zero-shot named entity recognition (NER) based on the jhu-clsp/mmBERT-base backbone (a ModernBERT-style multilingual encoder).

The model is trained jointly on English, German, Italian, French, and Spanish over roughly 21k distinct entity types, and is designed for prompted / label-list NER in real-world multilingual scenarios (documents, logs, PII detection, etc.).

As a baseline, we compare against urchade/gliner_multi-v2.1, which is built on microsoft/mdeberta-v3-base, a multilingual DeBERTa v3 model that extends BERT/RoBERTa with disentangled attention and an enhanced mask decoder, architectural features that improve token-level representations and benefit sequence-labeling tasks such as NER.

Demo

🔍 Demo: You can try the model directly in the browser via our Hugging Face Space:
https://huggingface.co/spaces/VAGOsolutions/SauerkrautLM-GLiNER-Demo

---

Model details

• Model name: SauerkrautLM-GLiNER
• Backbone: jhu-clsp/mmBERT-base (ModernBERT-style multilingual encoder)
• Architecture: GLiNER-style prompted NER (span classification conditioned on textual labels)
• Languages (training): English, German, French, Italian, Spanish
• Entity types: ~21k distinct entity labels
• Objective: Multilingual, zero-shot and few-shot NER with arbitrary label lists
• License: MIT

The model is trained as a general-purpose multilingual NER model: unlike a PII-only model, it is exposed to many domains (news, encyclopedic text, technical and general content) and a very large label space, making it suitable for both broad-domain extraction and specialized taxonomies.

Evaluation & Benchmarks

Datasets

The model is evaluated on the VAGOsolutions/gliner-benchmark-multilingual dataset (🤗 Dataset Link), which extends the original DFKI-SLT/cross_ner dataset (English-only) with German, French, Italian, and Spanish.

Important: The non-English data is not simply translated from the English original. Instead, new language-specific data was carefully curated to ensure high-quality, authentic entity extraction across all languages.

Results

We compare SauerkrautLM-GLiNER against two baselines:

• urchade/gliner_multi-v2.1 – General multilingual NER model
• urchade/gliner_multi_pii-v1 – Specialized PII detection model

CrossNER + Multilingual Domain Benchmarks (Threshold: 0.8 for SauerkrautLM-GLiNER, 0.65 for gliner_multi-v2.1)

Dataset	Samples	Entities	SauerkrautLM-GLiNER P / R / F1	gliner_multi-v2.1 P / R / F1
CrossNER (English)
crossner_ai_test.jsonl	431	14	58.03 / 48.92 / 53.09	64.94 / 40.24 / 49.69
crossner_literature_test.jsonl	416	12	55.57 / 47.79 / 51.39	73.11 / 51.59 / 60.49
crossner_music_test.jsonl	465	13	67.43 / 54.86 / 60.50	76.80 / 66.40 / 71.22
crossner_politics_test.jsonl	651	9	57.85 / 54.36 / 56.05	72.85 / 64.46 / 68.40
crossner_science_test.jsonl	543	17	57.83 / 47.01 / 51.86	67.05 / 54.10 / 59.88
AI Domain (Multilingual)
ai_french.jsonl	2311	20	53.76 / 66.13 / 59.30	56.68 / 22.17 / 31.88
ai_german.jsonl	2348	19	53.37 / 67.99 / 59.80	57.41 / 23.23 / 33.08
ai_italian.jsonl	2188	25	47.33 / 54.00 / 50.45	49.64 / 15.62 / 23.76
ai_spanish.jsonl	2253	19	54.23 / 63.96 / 58.69	50.09 / 21.63 / 30.21
Literature Domain (Multilingual)
literature_french.jsonl	541	17	45.64 / 59.36 / 51.60	36.75 / 18.16 / 24.31
literature_german.jsonl	505	20	34.59 / 41.94 / 37.92	30.13 / 10.04 / 15.07
literature_italian.jsonl	634	28	47.22 / 59.13 / 52.51	23.89 / 5.39 / 8.79
literature_spanish.jsonl	495	17	47.21 / 58.73 / 52.34	37.14 / 16.81 / 23.14
Music Domain (Multilingual)
music_french.jsonl	576	18	55.59 / 67.62 / 61.02	45.58 / 26.11 / 33.20
music_german.jsonl	457	18	53.68 / 65.01 / 58.81	54.95 / 31.94 / 40.40
music_italian.jsonl	555	23	71.75 / 73.07 / 72.40	45.48 / 16.92 / 24.66
music_spanish.jsonl	540	18	59.54 / 68.54 / 63.72	49.07 / 27.54 / 35.28
Politics Domain (Multilingual)
politics_french.jsonl	571	11	44.44 / 64.59 / 52.65	40.94 / 27.31 / 32.76
politics_german.jsonl	550	14	31.96 / 46.44 / 37.86	25.16 / 13.88 / 17.90
politics_italian.jsonl	574	17	49.64 / 68.87 / 57.70	41.52 / 21.04 / 27.93
politics_spanish.jsonl	559	13	50.86 / 69.71 / 58.81	46.14 / 29.71 / 36.15
Science Domain (Multilingual)
science_french.jsonl	563	25	59.24 / 61.42 / 60.31	55.55 / 13.89 / 22.23
science_german.jsonl	555	29	41.79 / 49.61 / 45.37	34.55 / 8.99 / 14.27
science_italian.jsonl	523	36	64.55 / 63.35 / 63.95	43.21 / 8.18 / 13.75
science_spanish.jsonl	555	32	54.95 / 55.95 / 55.45	23.89 / 5.88 / 9.43
AVERAGE	20,359	–	52.72 / 59.13 / 55.34	48.10 / 25.65 / 32.32

Key Takeaway: SauerkrautLM-GLiNER achieves +23.02 F1 points over gliner_multi-v2.1 on average, with particularly strong improvements in recall across all non-English languages.

---

PII Detection Benchmark (Threshold: 0.75 for SauerkrautLM-GLiNER, 0.6 for baselines)

Dataset	Samples	Entities	SauerkrautLM-GLiNER P / R / F1	gliner_multi-v2.1 P / R / F1	gliner_multi_pii-v1 P / R / F1
pii_english.jsonl	500	20	55.83 / 40.86 / 47.19	32.12 / 16.01 / 21.37	63.33 / 26.74 / 37.60
pii_french.jsonl	500	20	56.62 / 36.44 / 44.34	36.08 / 18.09 / 24.09	71.57 / 37.14 / 48.90
pii_german.jsonl	500	20	54.59 / 36.24 / 43.56	34.61 / 16.85 / 22.67	69.48 / 31.41 / 43.27
pii_italian.jsonl	500	20	58.14 / 38.01 / 45.97	39.80 / 19.93 / 26.56	71.40 / 34.38 / 46.41
pii_spanish.jsonl	500	20	55.05 / 36.13 / 43.62	39.86 / 20.24 / 26.85	66.84 / 33.25 / 44.41
AVERAGE	2,500	–	56.05 / 37.54 / 44.94	36.49 / 18.22 / 24.31	68.52 / 32.58 / 44.12

Key Takeaway: SauerkrautLM-GLiNER performs competitively on PII detection despite being a general-purpose NER model (not PII-specialized), achieving +20.63 F1 over gliner_multi-v2.1 and matching the specialized gliner_multi_pii-v1 model.

---

Usage

Installation

pip install gliner

Basic Example

from gliner import GLiNER

# Load the model
model = GLiNER.from_pretrained("VAGOsolutions/SauerkrautLM-GLiNER")

# Example text
text = """
Cristiano Ronaldo dos Santos Aveiro (Portuguese pronunciation: [kɾiʃˈtjɐnu ʁɔˈnaldu]; born 5 February 1985) is a Portuguese professional footballer who plays as a forward for and captains both Saudi Pro League club Al Nassr and the Portugal national team. Widely regarded as one of the greatest players of all time, Ronaldo has won five Ballon d'Or awards,[note 3] a record three UEFA Men's Player of the Year Awards, and four European Golden Shoes, the most by a European player. He has won 33 trophies in his career, including seven league titles, five UEFA Champions Leagues, the UEFA European Championship and the UEFA Nations League. Ronaldo holds the records for most appearances (183), goals (140) and assists (42) in the Champions League, goals in the European Championship (14), international goals (128) and international appearances (205). He is one of the few players to have made over 1,200 professional career appearances, the most by an outfield player, and has scored over 850 official senior career goals for club and country, making him the top goalscorer of all time.
"""

# Define labels (entity types you want to extract)
labels = ["person", "award", "date", "competitions", "teams"]

# Predict entities
entities = model.predict_entities(text, labels)

# Print results
for entity in entities:
    print(entity["text"], "=>", entity["label"])

Expected Output:

Cristiano Ronaldo dos Santos Aveiro => person
5 February 1985 => date
Al Nassr => teams
Portugal national team => teams
Ballon d'Or => award
UEFA Men's Player of the Year Awards => award
European Golden Shoes => award
UEFA Champions Leagues => competitions
UEFA European Championship => competitions
UEFA Nations League => competitions
Champions League => competitions
European Championship => competitions

Multilingual Example

The model works seamlessly across multiple languages:

# German text
text_de = "Angela Merkel war von 2005 bis 2021 Bundeskanzlerin der Bundesrepublik Deutschland."
labels_de = ["person", "position", "location", "date"]

entities_de = model.predict_entities(text_de, labels_de)
for entity in entities_de:
    print(entity["text"], "=>", entity["label"])

# French text
text_fr = "Emmanuel Macron est président de la République française depuis mai 2017."
labels_fr = ["person", "position", "location", "date"]

entities_fr = model.predict_entities(text_fr, labels_fr)
for entity in entities_fr:
    print(entity["text"], "=>", entity["label"])

Custom Labels for Domain-Specific NER

Simply adjust the label list to match your domain:

# PII Detection
text_pii = "Contact John Smith at [email protected] or call +1-555-0123."
labels_pii = ["person", "email", "phone"]

# Scientific Text
text_science = "The experiment used CRISPR-Cas9 to edit the BRCA1 gene in human cells."
labels_science = ["technique", "gene", "organism"]

# Business Documents
text_business = "Apple Inc. acquired startup.ai for $2.5 billion in Q3 2024."
labels_business = ["company", "money", "date"]

---

Intended use & limitations

Intended use

• Zero-shot or few-shot NER in:
  • Multilingual documents (EN/DE/FR/IT/ES)
  • Domain-specific content (AI, literature, music, politics, science)
  • PII / sensitive-data detection (emails, IDs, names, locations, etc.)
• Rapid prototyping of new NER schemas by simply changing the label list:
  • e.g. ["person", "organization", "product"]
  • or more domain-specific labels like ["award", "competition", "team", "date"]

Not intended for

• Safety-critical scenarios without human oversight (e.g. automated compliance decisions or legal / financial judgments).

---

Citation

If you use SauerkrautLM-GLiNER in your research or applications, please cite:

@misc{SauerkrautLM-GLiNER,
  title={SauerkrautLM-GLiNER},
  author={Michele Montebovi},
  organization={VAGO Solutions},
  url={https://huggingface.co/VAGOsolutions/SauerkrautLM-GLiNER},
  year={2025}
}

Acknowledgments

This model builds upon:

• mmBERT (jhu-clsp/mmBERT-base) – ModernBERT-style multilingual encoder
• GLiNER architecture (urchade/gliner_multi-v2.1) – Zero-shot NER framework