Brief-details: BERT-based hate speech classification model trained on Twitter/Gab data with human rationales. Detects hate speech, offensive content & normal text.
Brief-details: E5-small-v2 is a 12-layer text embedding model with 33.4M parameters, trained via weakly-supervised contrastive learning for semantic similarity tasks.
Brief-details: A self-supervised Vision Transformer (ViT) model for image feature extraction, trained on ImageNet-1k using DINO method, with small architecture and 16x16 patch size.
BRIEF DETAILS: RoBERTa-based model fine-tuned for irony detection in tweets, trained on 58M tweets. Achieves strong performance on TweetEval benchmark. Ideal for social media analysis.
BRIEF-DETAILS: Cross-encoder TinyBERT model trained on STS benchmark dataset for semantic similarity scoring, featuring efficient architecture and Apache 2.0 license.
Brief Details: DeiT-base is a data-efficient Vision Transformer with 86M parameters achieving 81.8% ImageNet accuracy, optimized for image classification tasks
Brief-details: A powerful 137M parameter English embedding model supporting 8192 token sequences, using ALiBi positioning for enhanced long-text processing and RAG applications.
Brief Details: Vikhr-7B-instruct_0.4 is a bilingual Russian-English LLM with 7.63B parameters, featuring enhanced SFT training and improved JSON/multiturn capabilities.
Brief-details: A 40B parameter instruction-tuned LLM from TII, based on Falcon-40B. Optimized for chat/instruction tasks with FlashAttention and multiquery architecture.
Brief-details: A 1.5B parameter sentence embedding model trained on GTE-large with multi-dimensional retrieval learning, optimized for semantic search and text similarity tasks
Brief Details: HuBERT base model for self-supervised speech representation learning, trained on LibriSpeech. Features 16kHz audio processing and BERT-like prediction architecture.
Brief Details: DINOv2-giant: A 1.14B parameter Vision Transformer model for self-supervised image feature extraction, developed by Facebook for robust visual understanding.
BRIEF DETAILS: DeBERTa V2 XLarge: Advanced NLP model with 900M parameters, featuring disentangled attention and enhanced mask decoder. Achieves SOTA performance on NLU tasks.
Brief-details: Italian BERT model trained on 81GB corpus with 13B tokens. Uncased version with 111M parameters optimized for Italian language tasks.
Brief Details: Japanese text embedding model (133M params) optimized for retrieval tasks with SOTA performance, supports CPU inference and cosine similarity.
Brief-details: Efficient zero-shot text classifier based on DeBERTa-v3-large (435M params), trained for universal classification tasks with commercial-friendly data and strong performance.
Brief-details: F-coref is a fast and accurate coreference resolution model achieving 78.5 F1 on OntoNotes, processing 2.8K documents in 25 seconds on V100 GPU.
BRIEF-DETAILS: Advanced 70B parameter LLaMA-2 variant with extended 32k token context window, optimized for long-form content processing and generation
Brief Details: A Helsinki-NLP English-to-Russian neural machine translation model with strong BLEU scores (31.1 on newstest2012), based on the Marian framework
Brief-details: Optimized 765M parameter Llama 3.2 model with 4-bit quantization for efficient inference. Supports multilingual tasks with reduced memory footprint and faster processing.
Brief-details: DeBERTa-v3 model fine-tuned for prompt injection detection, achieving 99.14% accuracy. Built by deepset, it classifies requests as INJECTION or LEGIT.
