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.
As a llama-based compact model, Prompt-Guard-86M focuses on text classification. Weighing in near 86M parameters, Prompt-Guard-86M trades some ceiling for cheaper, faster inference. Prompt-Guard-86M is subject to Llama 3.1 Community terms, so confirm licensing before commercial use. Prompt-Guard-86M ships without a hosted SLA, so budget for self-managed deployment and monitoring.
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.
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.
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.
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.
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.
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.
As a bert-based open-weight model, bert-base-multilingual-uncased-sentiment focuses on text classification. Training spans multiple languages, so bert-base-multilingual-uncased-sentiment covers cross-lingual text classification from one checkpoint. The MIT license keeps bert-base-multilingual-uncased-sentiment unrestricted for commercial reuse. Before relying on bert-base-multilingual-uncased-sentiment, reproduce its key numbers on representative inputs.
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.
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.
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.
Built for text classification, finbert-tone is a bert-based model with publicly available weights. Read finbert-tone's card for hardware requirements and licensing fine print before deploying.
turn-detector is an open-weight text classification model in the llama family. turn-detector is multilingual by design rather than English-only. Licensing for turn-detector is unspecified or custom — clear it before commercial use. turn-detector is community-maintained, so track upstream changes and pin a known-good revision.
As a distilbert-based open-weight model, distilbert-base-multilingual-cased-sentiments-student focuses on text classification. Training spans multiple languages, so distilbert-base-multilingual-cased-sentiments-student covers cross-lingual text classification from one checkpoint. The Apache 2.0 license keeps distilbert-base-multilingual-cased-sentiments-student unrestricted for commercial reuse. Before relying on distilbert-base-multilingual-cased-sentiments-student, reproduce its key numbers on representative inputs.
jina-reranker-m0 is a multimodal reranker fine-tuned from Qwen2-VL-2B-Instruct, capable of scoring text and image relevance for retrieval pipelines. It is tagged for the Vidore benchmark, indicating focus on document and visual retrieval reranking. The CC-BY-NC-4.0 license restricts commercial use.
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.
robertuito-sentiment-analysis classifies text into predefined label categories using a RoBERTa encoder fine-tuned with a classification head. It outputs per-class logits.
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.
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.
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.
CryptoBERT is a RoBERTa-based sentiment classifier fine-tuned on the StockTwits crypto dataset, designed specifically for cryptocurrency-related social media text. It outputs three-class sentiment labels (bullish, bearish, neutral) and targets the linguistic patterns common in crypto discourse on platforms like Twitter and StockTwits. The model is trained exclusively on English-language crypto content, making it narrow in scope but well-suited for that domain.
As a bert-based open-weight model, KR-FinBert-SC focuses on text classification. KR-FinBert-SC ships without a hosted SLA, so budget for self-managed deployment and monitoring.
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.
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.
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-weight model aimed at text classification. internlm2-1_8b-reward's 8000M-parameter size keeps hosting requirements modest relative to frontier models. internlm2-1_8b-reward lists a non-standard license, so confirm permissions before deployment. internlm2-1_8b-reward ships without a hosted SLA, so budget for self-managed deployment and monitoring.
deberta-v3-base-prompt-injection-v2 targets text classification and is shipped as an open-weight, self-hostable checkpoint. It is a fine-tune of deberta-v3-base, inheriting that base model's general competence. Permissive Apache 2.0 terms let deberta-v3-base-prompt-injection-v2 go straight into commercial pipelines. Treat deberta-v3-base-prompt-injection-v2's published metrics as a starting point and validate against your workload.
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.
Built for text classification, xlm-emo-t is a xlm roberta-based model with publicly available weights. Training spans multiple languages, so xlm-emo-t covers cross-lingual text classification from one checkpoint. Check the xlm-emo-t model card for benchmarks and intended use before adopting it.
phishing-email-detection-distilbert_v2.4.1 is an open-weight checkpoint for text classification, distributed on the HuggingFace Hub. The Apache 2.0 license keeps phishing-email-detection-distilbert_v2.4.1 unrestricted for commercial reuse. It is a fine-tune of distilbert-base-uncased, inheriting that base model's general competence. Like most open checkpoints, phishing-email-detection-distilbert_v2.4.1 rewards a quick in-domain eval before commitment.
inclusively-classification is a bert-based open-weight model aimed at text classification. Because inclusively-classification uses CC BY-NC-SA 4.0, vet the conditions against your deployment plan. Check the inclusively-classification model card for benchmarks and intended use before adopting it.
distilroberta-finetuned-financial-news-sentiment-analysis is an openly licensed text classification model in the roberta family. distilroberta-finetuned-financial-news-sentiment-analysis is Apache 2.0-licensed, clearing it for closed-source and paid products. Evaluate distilroberta-finetuned-financial-news-sentiment-analysis on your own data before trusting it in production.
distilbert-base-uncased-emotion is an open-weight checkpoint for text classification, distributed on the HuggingFace Hub. The Apache 2.0 license keeps distilbert-base-uncased-emotion unrestricted for commercial reuse. Like most open checkpoints, distilbert-base-uncased-emotion rewards a quick in-domain eval before commitment.
rubert-base-cased-sentiment-rusentiment is an open-weight text classification model in the bert family. Evaluate rubert-base-cased-sentiment-rusentiment on your own data before trusting it in production.
Built for text classification, phobert-base-vietnamese-sentiment is a roberta-based model with publicly available weights. phobert-base-vietnamese-sentiment is MIT-licensed, clearing it for closed-source and paid products. Check the phobert-base-vietnamese-sentiment model card for benchmarks and intended use before adopting it.
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 openly licensed text classification model in the bert family. FinBERT-PT-BR is Apache 2.0-licensed, clearing it for closed-source and paid products. Treat FinBERT-PT-BR's published metrics as a starting point and validate against your workload.
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-weight checkpoint for text classification, distributed on the HuggingFace Hub. Treat twitter-xlm-roberta-base-sentiment-multilingual's published metrics as a starting point and validate against your workload.
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.