47 models · ranked by HuggingFace downloads
BGE-Reranker-v2-M3 is BAAI's multilingual cross-encoder reranker built on XLM-RoBERTa, designed for re-ranking retrieved passages in multilingual RAG or search pipelines. It jointly encodes query-passage pairs to produce relevance scores, providing higher accuracy than bi-encoder similarity for the same candidate set. Apache 2.0 licensed with text-embeddings-inference support.
FinBERT is a BERT model fine-tuned on financial news and financial communications text for financial sentiment analysis, classifying text as positive, negative, or neutral from a finance domain perspective. Developed by Prosus AI (Naspers), it targets applications where general-purpose sentiment models fail on financial jargon and market-specific framing.
Cross-encoder reranker trained on multilingual data (English and Chinese) using XLM-RoBERTa. It scores query-document pairs directly rather than comparing embeddings, making it more accurate than bi-encoders for retrieval pipelines at the cost of higher latency.
RoBERTa-base fine-tuned on ~124M tweets for three-class sentiment classification (positive/neutral/negative). Trained by Cardiff NLP on the TweetEval benchmark, it consistently ranks among the top-performing tweet-specific sentiment models.
distilbert-base-uncased-finetuned-sst-2-english is a DistilBERT model fine-tuned on the Stanford Sentiment Treebank v2 (SST-2) for binary positive/negative sentiment classification of English text. It is one of the most downloaded sentiment classifiers on HuggingFace and commonly used as a demonstration or fast baseline. The model achieves approximately 91% accuracy on the SST-2 validation set.
twitter-xlm-roberta-base-sentiment classifies text into predefined label categories using a RoBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
RADAR-Vicuna-7B maps input sequences to one or more labels. Fine-tuned on labeled data, it covers tasks like sentiment analysis, topic detection, and intent classification.
Prompt-Guard-86M is a sequence classifier built on a DeBERTa backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
This RoBERTa-base model fine-tuned on Google's GoEmotions dataset classifies English text into 28 emotion categories or groups them into sentiment buckets. It supports multi-label classification, meaning a sentence can be tagged with several emotions simultaneously. The model is widely used in sentiment analytics pipelines and social media monitoring tools.
tiny-Qwen2ForSequenceClassification-2.5 maps input sequences to one or more labels. Fine-tuned on labeled data, it covers tasks like sentiment analysis, topic detection, and intent classification.
roberta-base-openai-detector is a RoBERTa-base binary classifier trained to distinguish GPT-2-generated text from human-written text. It was released by the OpenAI Grover team and works by fine-tuning on paired human/machine samples from the GPT-2 output corpus. As an early-generation AI text detector, it is most accurate on GPT-2 output and significantly less reliable on newer LLMs.
bert-base-multilingual-uncased-sentiment is a sequence classifier built on a BERT backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
twitter-roberta-base-sentiment maps input sequences to one or more labels. Fine-tuned on labeled data, it covers tasks like sentiment analysis, topic detection, and intent classification.
emotion-english-distilroberta-base classifies text into predefined label categories using a RoBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
finbert-tone is a sequence classifier built on a BERT backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
sentiment-polish-gpt2-small is a GPT2-small fine-tuned for Polish sentiment classification, trained on the PolEmo 2.0 dataset with four sentiment classes (positive, negative, ambivalent, neutral). GPT2 architecture fine-tunes cleanly for sequence classification despite being a causal LM by adding a classification head. It is one of the few openly available Polish-specific sentiment models at small scale.
turn-detector is a sequence classifier built on a Llama backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
distilbert-base-multilingual-cased-sentiments-student is a sequence classifier built on a DistilBERT backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
bge-reranker-v2-gemma is BAAI's cross-encoder reranker built on a Gemma backbone, trained for multilingual passage reranking. Unlike bi-encoders that embed query and document independently, it attends jointly to both as a cross-encoder for higher reranking accuracy. It is designed to be used as the second stage after a bi-encoder retrieval step in two-stage RAG pipelines.
xlm-roberta-base-language-detection maps input sequences to one or more labels. Fine-tuned on labeled data, it covers tasks like sentiment analysis, topic detection, and intent classification.
jina-reranker-m0 is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
robertuito-sentiment-analysis classifies text into predefined label categories using a RoBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
KoELECTRA-small-v3 fine-tuned on the Naver Sentiment Movie Corpus (NSMC) for Korean binary sentiment classification. It uses the ELECTRA discriminator architecture, which is more parameter-efficient than BERT for the same task performance.
Qwen2.5-1.5B-apeach is a Korean hate speech detection classifier fine-tuned on the APEACH dataset using Qwen2.5-1.5B as the backbone. APEACH (Automated Pipeline for Evaluation Against Crowdsourced Hate speech) is a Korean benchmark for detecting offensive and hateful content. This model converts the LLM into a sequence classifier for binary or multi-class Korean content moderation.
MedCPT-Cross-Encoder is a BERT-based cross-encoder from NCBI fine-tuned for medical text relevance scoring, trained on PubMed query-article pairs. It takes a query and a candidate passage and scores their relevance, making it a reranking component in medical information retrieval pipelines. The model is typically paired with MedCPT's query encoder for a full retrieval-reranking system.
Meta's fastText-based language identification model, capable of identifying 176 languages from short text strings. Extremely fast CPU inference makes it practical for preprocessing pipelines that need to route text by language.
A W4A16 (4-bit weights, 16-bit activations) quantized version of the Qwen3-Reranker-4B, enabling efficient cross-encoder reranking inference. The group-size-128 quantization balances compression ratio against accuracy retention for reranking tasks.
KR-FinBert-SC is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
A 0.6B cross-encoder reranker from farbodtavakkoli specialized for reranking OpenTelemetry log, trace, and metric documents. Pairs with OTel-Embedding models for a two-stage observability retrieval pipeline.
RoBERTa-large fine-tuned on the Multi-Genre Natural Language Inference (MNLI) corpus, commonly used for zero-shot text classification via the NLI entailment trick. One of the most frequently used models for zero-shot classification before dedicated models like DeBERTa-MNLI improved further.
cryptobert is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
deberta-large-mnli classifies text into predefined label categories using a DeBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
internlm2-1_8b-reward is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
deberta-v3-base-prompt-injection-v2 is a sequence classifier built on a DeBERTa backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
BETO-based sentiment analysis model from finiteautomata, fine-tuned for Spanish sentiment classification. BETO is the Spanish BERT model, and this checkpoint targets positive/negative/neutral classification on Spanish text.
xlm-emo-t is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
phishing-email-detection-distilbert_v2.4.1 is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
inclusively-classification is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
distilroberta-finetuned-financial-news-sentiment-analysis is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
distilbert-base-uncased-emotion is a sequence classifier built on a DistilBERT backbone. Given a string, it scores each candidate label and returns the highest-confidence prediction.
rubert-base-cased-sentiment-rusentiment is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
phobert-base-vietnamese-sentiment is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
multilingual-sentiment-analysis classifies text into predefined label categories using a DistilBERT encoder fine-tuned with a classification head. It outputs per-class logits.
FinBERT-PT-BR is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
roberta_toxicity_classifier classifies text into predefined label categories using a RoBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
twitter-xlm-roberta-base-sentiment-multilingual is an open-source text-classification model available on HuggingFace. Details are sourced from the public model registry.
DeBERTa-XLarge fine-tuned on the Multi-Genre Natural Language Inference (MNLI) dataset for zero-shot text classification via entailment. DeBERTa's disentangled attention mechanism improves on BERT for NLI tasks. The xlarge variant (900M parameters) provides strong NLI accuracy but is expensive for high-throughput use.