25 models · ranked by HuggingFace downloads
bert-large-cased-finetuned-conll03-english assigns labels to individual tokens in a sequence, directly applicable to named entity recognition, part-of-speech tagging, and span extraction.
indonesian-roberta-base-posp-tagger performs sequence labeling: each input token receives a class label aligned to its text position. Typical tasks include NER, chunking, and slot filling.
bert-base-NER performs sequence labeling: each input token receives a class label aligned to its text position. Typical tasks include NER, chunking, and slot filling.
stanford-deidentifier-base uses a BERT encoder with a per-token classification head. The BIO tagging scheme is standard for its NER fine-tunes.
GGUF-quantized privacy filter built on OpenMed's privacy-filter-nemotron, trained on NVIDIA's Nemotron-PII dataset. Identifies and redacts 10+ PII categories in English text including names, email addresses, and medical identifiers. Designed for LocalAI's privacy-filter.cpp runtime.
A DistilBERT model fine-tuned on Twitter/X data for token classification tasks, likely part-of-speech tagging or named entity recognition on social media text. BERTweet-based initialization means it handles informal spelling, hashtags, and abbreviations better than standard BERT. The training split and label schema are not publicly documented.
KoELECTRA-small-v3 fine-tuned on the Modu corpus for Korean named entity recognition across nine standard NER categories. Uses ELECTRA's discriminator architecture for efficient token-level classification without generative overhead. Targets CPU-deployable Korean NLP pipelines.
punctuate-all uses a RoBERTa encoder with a per-token classification head. The BIO tagging scheme is standard for its NER fine-tunes.
Built for token classification and NER, bert-portuguese-ner is a bert-based model with publicly available weights. bert-portuguese-ner is MIT-licensed, clearing it for closed-source and paid products. Check the bert-portuguese-ner model card for benchmarks and intended use before adopting it.
SAT-3l-sm (Segment Any Text, 3-layer small) is a multilingual text segmentation model supporting over 85 languages, designed to split continuous text into meaningful sentence or paragraph segments. Unlike rule-based sentence tokenisers that rely on punctuation, SAT uses a contextual XLM-based token classifier to handle languages with unusual or absent punctuation conventions. The small variant trades some accuracy for faster inference.
wikineural-multilingual-ner performs sequence labeling: each input token receives a class label aligned to its text position. Typical tasks include NER, chunking, and slot filling.
deid_roberta_i2b2 is a RoBERTa model fine-tuned on the i2b2 de-identification dataset to detect and classify protected health information (PHI) in clinical notes. It identifies PHI spans such as names, dates, locations, and IDs. MIT-licensed for integration into clinical NLP de-identification pipelines.
llmlingua-2-xlm-roberta-large-meetingbank assigns labels to individual tokens in a sequence, directly applicable to named entity recognition, part-of-speech tagging, and span extraction.
xlm-roberta-large-ner-hrl uses a RoBERTa encoder with a per-token classification head. The BIO tagging scheme is standard for its NER fine-tunes.
Microsoft's LLMLingua-2 token classification model trained on MeetingBank data for prompt compression. It classifies tokens as retain or drop, enabling lossless or near-lossless compression of long context inputs before passing them to an LLM.
Flair's fast English NER model using the Flair framework's sequence labeling approach with character-level language model embeddings. 'Fast' indicates a smaller, speed-optimized variant compared to Flair's standard NER model. Recognizes standard NE classes (PER, ORG, LOC, MISC).
Built for token classification and NER, privacy-filter is a model with publicly available weights. privacy-filter is Apache 2.0-licensed, clearing it for closed-source and paid products. Check the privacy-filter model card for benchmarks and intended use before adopting it.
XLM-RoBERTa-base fine-tuned for named entity recognition across high-resource languages (HRL). The model handles person, organization, and location entity extraction in a multilingual setting using token classification. XLM-RoBERTa's multilingual pretraining makes it competitive on languages it did not see explicit NER supervision for.
layoutreader is a LayoutLMv3-based token classification model designed to predict reading order in document layouts. Built on top of Microsoft's LayoutLMv3, it classifies tokens by their sequential reading position, making it useful for document parsing pipelines that need to linearize visual document structure.
BERT-small fine-tuned by Gravitee.io for PII detection via token classification, identifying entities such as names, email addresses, and phone numbers. Built on prajjwal1/bert-small (~29M parameters), designed for integration into API gateway or data-pipeline filtering logic where inference latency is critical.
A BERT-base model fine-tuned for named entity recognition on Russian text. Handles standard NER categories (persons, organizations, locations) on Russian-language inputs using the standard token-classification approach.
fullstop-punctuation-multilang-large assigns labels to individual tokens in a sequence, directly applicable to named entity recognition, part-of-speech tagging, and span extraction.
roberta-large-ner-english is a roberta-based open-weight model aimed at token classification and NER. Permissive MIT terms let roberta-large-ner-english go straight into commercial pipelines. Read roberta-large-ner-english's card for hardware requirements and licensing fine print before deploying.
bert-base-chinese-ws is an open-weight token classification and NER model in the bert family. Distribution of bert-base-chinese-ws is under GPL-3.0, which is worth reading before you ship. Like most open checkpoints, bert-base-chinese-ws rewards a quick in-domain eval before commitment.
As a bert-based open-weight model, bert-base-multilingual-cased-ner-hrl focuses on token classification and NER. bert-base-multilingual-cased-ner-hrl lists a non-standard license, so confirm permissions before deployment. Check the bert-base-multilingual-cased-ner-hrl model card for benchmarks and intended use before adopting it.