knowledgator/gliclass-multilang-mini

GLiClass Multilang: Efficient multilingual zero-shot and few-shot multi-task model via sequence classification

GLiClass is an efficient zero-shot sequence classification model designed to achieve SoTA performance while being much faster than cross-encoders and LLMs, while preserving strong generalization capabilities.

The model supports text classification with any labels and can be used for the following tasks:

• Topic Classification
• Sentiment Analysis
• Intent Classification
• Reranking
• Hallucination Detection
• Rule-following Verification
• LLM-safety Classification
• Natural Language Inference

✨ What's New in GLiClass Multilang

• Multilingual Training — Natively trained on 20 languages: Swedish, Norwegian, Czech, Polish, Lithuanian, Estonian, Latvian, Spanish, Finnish, German, French, Romanian, Italian, Portuguese, Dutch, Ukrainian, Hindi, Chinese, Arabic, and Hebrew.
• Cross-lingual Classification — Labels and input texts can be in different languages; classify a German document with English labels, or mix languages freely across inputs and labels.
• CrossAttn Scorer — A new cross-attention scorer enables more efficient pooling independently for each label with unpadding and flash-attn.
• Hierarchical Labels — Organize labels into groups using dot notation or dictionaries (e.g., sentiment.positive, topic.product).
• Few-Shot Examples — Provide in-context examples to boost accuracy on your specific task.
• Label Descriptions — Add natural-language descriptions to labels for more precise classification.
• Task Prompts — Prepend a custom prompt to guide the model's classification behavior.

See the GLiClass library README for full details on these features.

Installation

pip install gliclass

Quick Start

from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer

model = GLiClassModel.from_pretrained("knowledgator/gliclass-multilang-mini")
tokenizer = AutoTokenizer.from_pretrained("knowledgator/gliclass-multilang-mini")
pipeline = ZeroShotClassificationPipeline(model, tokenizer, classification_type='multi-label', device='cuda:0')

text = "NASA launched a new Mars rover to search for signs of ancient life."
labels = ["space", "politics", "sports", "technology", "health"]

results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

---

Multilingual & Cross-lingual Capabilities

Natively trained on 20 languages. Labels and texts can be in different languages.

Same language (German):

from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer

model = GLiClassModel.from_pretrained("knowledgator/gliclass-multilang-mini")
tokenizer = AutoTokenizer.from_pretrained("knowledgator/gliclass-multilang-mini")
pipeline = ZeroShotClassificationPipeline(model, tokenizer, classification_type='multi-label', device='cuda:0')

text = "Die NASA hat einen neuen Mars-Rover gestartet, um nach Spuren alten Lebens zu suchen."
labels = ["Weltraum", "Politik", "Sport", "Technologie", "Gesundheit"]
results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

Cross-lingual (French text, English labels):

text = "Le gouvernement français a annoncé de nouvelles mesures économiques."
labels = ["economy", "politics", "sports", "technology"]
results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

Cross-lingual (Arabic text, English labels):

text = "أطلقت ناسا مركبة جديدة للمريخ للبحث عن آثار الحياة القديمة."
labels = ["space", "politics", "sports", "technology"]
results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

Cross-lingual (English text, Spanish labels):

text = "NASA launched a new Mars rover to search for signs of ancient life."
labels = ["espacio", "política", "deportes", "tecnología", "salud"]
results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

General Examples

1. Topic Classification

text = "NASA launched a new Mars rover to search for signs of ancient life."
labels = ["space", "politics", "sports", "technology", "health"]

results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

With hierarchical labels

hierarchical_labels = {
    "science": ["space", "biology", "physics"],
    "society": ["politics", "economics", "culture"]
}

results = pipeline(text, hierarchical_labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])
# e.g. science.space => 0.95

2. Sentiment Analysis

text = "The food was excellent but the service was painfully slow."
labels = ["positive", "negative", "neutral"]

results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

With a task prompt

results = pipeline(
    text, labels,
    prompt="Classify the sentiment of this restaurant review:",
    threshold=0.5
)[0]

3. Intent Classification

text = "Can you set an alarm for 7am tomorrow?"
labels = ["set_alarm", "play_music", "get_weather", "send_message", "set_reminder"]

results = pipeline(text, labels, threshold=0.5)[0]
for r in results:
    print(r["label"], "=>", r["score"])

4. Natural Language Inference

Represent your premise as the text and the hypothesis as a label. The model works best with a single hypothesis at a time.

text = "The cat slept on the windowsill all afternoon."
labels = ["The cat was awake and playing outside."]

results = pipeline(text, labels, threshold=0.0)[0]
print(results)
# Low score → contradiction

5. Reranking

Score query–passage relevance by treating passages as texts and the query as the label:

query = "How to train a neural network?"
passages = [
    "Backpropagation is the key algorithm for training deep neural networks.",
    "The stock market rallied on strong earnings reports.",
    "Gradient descent optimizes model weights during training.",
]

for passage in passages:
    score = pipeline(passage, [query], threshold=0.0)[0][0]["score"]
    print(f"{score:.3f}  {passage[:60]}")

6. Rule-following Verification

Include the domain and rules as part of the text:

text = (
    "Domain: e-commerce product reviews\n"
    "Rule: No promotion of illegal activity.\n"
    "Text: The software is okay, but search for 'productname_patch_v2.zip' "
    "to unlock all features for free."
)
labels = ["follows_guidelines", "violates_guidelines"]

results = pipeline(text, labels, threshold=0.0)[0]
for r in results:
    print(r["label"], "=>", r["score"])

---

Benchmarks

Model Overview

Summary across all evaluated multilingual-capable models (zero-shot, no fine-tuning). Speed averaged over all label counts and text lengths at batch_size=8 on NVIDIA RTX PRO 6000 Blackwell.

Model	Params	English avg F1	Multilingual avg F1	Throughput (samp/s, bs=8)
multilang‑ultra	~1 720M	0.7212	0.5599	200.7
multilang‑mini	~288M	0.6827	0.5378	513.4
multilang‑edge	~140M	0.6196	0.3959	553.6
instruct‑large	~435M	0.7199	—	293.9
instruct‑base	~184M	0.6525	—	521.9
gliner2‑large‑v1	340M	0.6774	—	122.5
gliner2‑multi‑v1	~278M	0.6387	0.4659	200.2
gliner2‑base‑v1	~184M	0.6336	—	224.0
bge‑m3‑zeroshot‑v2.0	568M	0.5927	0.5225	208.7
mDeBERTa‑mnli	300M	0.5340	0.3926	160.6

Multilingual avg F1 is the mean of 6 dataset-level scores (GermEval2017, MASSIVE, PolygloToxicityPrompts, SIB-200, TextDetox, TweetSentiment). Models without multilingual results (—) were only evaluated on English datasets.

---

F1 scores on zero-shot text classification (no fine-tuning on these datasets):

Table A: GLiClass Multilang (macro F1)

Dataset	multilang‑ultra	multilang‑mini	multilang‑edge
CR	0.9226	0.9042	0.8852
sst2	0.9065	0.8810	0.8276
sst5	0.3049	0.2806	0.3047
20_newsgroups	0.5238	0.4242	0.3522
spam	0.9625	0.9385	0.6787
financial_phrasebank	0.8724	0.7156	0.7446
imdb	0.9330	0.9011	0.8730
ag_news	0.7454	0.7545	0.7338
emotion	0.4825	0.4655	0.4267
cap_sotu	0.4385	0.4087	0.3516
rotten_tomatoes	0.8413	0.8236	0.7044
massive	0.6483	0.5853	0.5649
banking	0.6492	0.5853	0.5788
snips	0.8653	0.8900	0.6487
AVERAGE	0.7212	0.6827	0.6196

Table B: Baselines (macro F1)

Dataset	gliner2‑large‑v1	gliner2‑multi‑v1	gliner2‑base‑v1	bge‑m3‑zeroshot‑v2.0	mDeBERTa‑mnli
CR	0.9117	0.8785	0.8783	0.9041	0.8956
sst2	0.8911	0.8568	0.8737	0.9257	0.8516
sst5	0.4462	0.3784	0.4100	0.2931	0.3023
20_newsgroups	0.5163	0.3668	0.4608	0.4161	0.2080
spam	0.3558	0.5986	0.3843	0.4410	0.4980
financial_phrasebank	0.8330	0.7372	0.7225	0.5040	0.4444
imdb	0.9170	0.8934	0.8982	0.8730	0.8264
ag_news	0.7029	0.7403	0.7193	0.6870	0.6547
emotion	0.5233	0.4666	0.4577	0.4530	0.4055
cap_sotu	0.4387	0.3972	0.3831	0.4720	0.3390
rotten_tomatoes	0.7909	0.7210	0.6979	0.8130	0.6931
massive	0.5897	0.4721	0.5403	0.4140	0.2527
banking	0.6885	0.6390	0.6709	0.3870	0.3796
snips	0.8788	0.7954	0.7731	0.7149	0.7245
AVERAGE	0.6774	0.6387	0.6336	0.5927	0.5340

Table C: GLiClass-V1 Multitask (macro F1)

Dataset	instruct‑large‑v1.0	instruct‑base‑v1.0	edge‑v1.0
CR	0.9066	0.8922	0.7933
sst2	0.9154	0.9198	0.7577
sst5	0.3387	0.2266	0.2163
20_newsgroups	0.5577	0.5189	0.2555
spam	0.9790	0.9380	0.7609
financial_phrasebank	0.8289	0.5217	0.3905
imdb	0.9397	0.9364	0.8159
ag_news	0.7521	0.6978	0.6043
emotion	0.4473	0.4454	0.2941
cap_sotu	0.4327	0.4579	0.2380
rotten_tomatoes	0.8491	0.8458	0.5455
massive	0.5824	0.4757	0.2090
banking	0.6987	0.6072	0.4635
snips	0.8509	0.6515	0.5461
AVERAGE	0.7199	0.6525	0.4922

Multilingual Benchmarks

Macro F1 averaged per dataset across all evaluated languages:

Dataset	multilang‑ultra	multilang‑mini	multilang‑edge	gliner2‑multi‑v1	bge‑m3‑zeroshot‑v2.0	mDeBERTa‑mnli
germeval2017	0.4647	0.4826	0.4094	0.4223	0.4503	0.2849
massive	0.5635	0.4925	0.2853	0.3625	0.4646	0.2427
polyglot_toxicity	0.7367	0.7110	0.4474	0.6630	0.6809	0.5698
sib200	0.1935	0.1921	0.1492	0.1750	0.1891	0.1476
textdetox	0.7428	0.7313	0.5811	0.5912	0.7510	0.6490
tweet_sentiment	0.6579	0.6171	0.5030	0.5814	0.5991	0.4615
AVERAGE	0.5599	0.5378	0.3959	0.4659	0.5225	0.3926

Per-language macro F1 (16-language fair comparison on massive + sib200):

Language	multilang‑ultra	multilang‑mini	multilang‑edge	gliner2‑multi‑v1	bge‑m3‑zeroshot‑v2.0	mDeBERTa‑mnli
arabic	0.3210	0.3043	0.1843	0.2394	0.2862	0.1567
chinese	0.3888	0.3636	0.2724	0.2947	0.3459	0.2356
dutch	0.3949	0.3587	0.2660	0.2828	0.3284	0.2146
finnish	0.3632	0.3174	0.1172	0.2704	0.3357	0.1884
french	0.3965	0.3679	0.2963	0.2946	0.3396	0.1978
german	0.3654	0.3457	0.2532	0.2767	0.3164	0.1966
hebrew	0.3521	0.3206	0.1271	0.2641	0.3287	0.1796
hindi	0.3934	0.3529	0.1877	0.0817	0.3240	0.1986
italian	0.3919	0.3474	0.2604	0.2891	0.3146	0.1976
latvian	0.3643	0.3165	0.1205	0.2741	0.3163	0.1774
norwegian	0.3770	0.3489	0.2043	0.2803	0.3382	0.1965
polish	0.3961	0.3577	0.2112	0.2814	0.3225	0.1981
portuguese	0.4008	0.3482	0.2798	0.3057	0.3346	0.1936
romanian	0.3740	0.3204	0.2210	0.2831	0.3291	0.1944
spanish	0.3921	0.3535	0.2905	0.2924	0.3371	0.1918
swedish	0.3863	0.3547	0.2121	0.2799	0.3317	0.2019
AVERAGE	0.3786	0.3424	0.2190	0.2681	0.3268	0.1950

Throughput

Throughput (samples/sec), batch_size=8, GPU: NVIDIA RTX PRO 6000 Blackwell. Averaged over text lengths (64 / 256 / 512 tokens).

Model	1 label	2	4	8	16	32	64	128	256	avg
multilang‑ultra	308.2	302.5	281.8	266.3	235.9	190.5	125.2	64.7	31.5	200.7
multilang‑mini	708.4	703.9	692.5	664.2	618.1	518.1	396.1	221.2	98.2	513.4
multilang‑edge	697.0	699.7	689.5	671.0	637.7	553.3	469.8	345.2	219.2	553.6
instruct‑large	397.2	393.1	386.6	374.2	351.1	313.3	223.8	142.2	63.2	293.9
instruct‑base	708.0	707.5	693.5	666.4	616.7	526.5	405.5	248.1	124.9	521.9
gliner2‑large‑v1	165.6	165.2	157.1	155.6	142.1	122.1	98.6	65.6	31.0	122.5
gliner2‑multi‑v1	270.4	267.9	264.6	257.3	237.2	200.0	159.2	96.8	48.4	200.2
gliner2‑base‑v1	296.8	293.2	287.8	278.9	262.0	229.4	180.1	121.3	66.2	224.0
bge‑m3‑zeroshot‑v2.0	940.0	474.7	238.4	112.9	58.3	28.9	14.4	7.2	3.7	208.7
mDeBERTa‑mnli	717.5	364.5	183.1	91.8	45.7	22.8	11.4	5.7	3.0	160.6

NLI models (bge-m3, mDeBERTa) run one forward pass per label — throughput drops linearly with label count. GLiClass and GLiNER2 encode all labels in a single pass, so throughput stays nearly flat.

Citation

@misc{stepanov2025gliclassgeneralistlightweightmodel,
      title={GLiClass: Generalist Lightweight Model for Sequence Classification Tasks}, 
      author={Ihor Stepanov and Mykhailo Shtopko and Dmytro Vodianytskyi and Oleksandr Lukashov and Alexander Yavorskyi and Mykyta Yaroshenko},
      year={2025},
      eprint={2508.07662},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.07662}, 
}