Brief Details: Wizard-Vicuna-13B is a fine-tuned 13B parameter LLM combining Vicuna's capabilities with wizard learning, optimized for instruction-following and dialogue
Brief Details: WizardLM-7B-GPTQ is a quantized version of WizardLM, optimized for efficiency with 4-bit precision and 128 group size, making it more accessible for deployment.
Brief-details: BERT-mini based encoder-decoder model fine-tuned on CNN/DailyMail dataset for text summarization, achieving 16.51 ROUGE-2 score. Lightweight and efficient.
Brief-details: Italian BERT model fine-tuned on SQuAD-it dataset for Q&A tasks, achieving 74.16% F1 score. Trained on 13GB Italian corpus with 2B+ tokens.
Brief-details: French sequence-to-sequence model based on BART architecture, trained on 66GB text data. Features 458M parameters for advanced NLP tasks like summarization.
BRIEF-DETAILS: French sequence-to-sequence model based on BART architecture, with 165M/458M params, specialized in generative tasks like summarization
BRIEF DETAILS: Tamil language model based on ELECTRA architecture, trained on 11.5GB corpus. Outperforms mBERT in news classification (75.1% vs 53%) and sentiment analysis.
Brief-details: DistilKoBERT is a lightweight Korean BERT model created by monologg, optimized for efficiency while maintaining Korean language understanding capabilities
Brief-details: Arabic speech recognition model fine-tuned on Wav2Vec2-Large-XLSR-53 achieving 36.69% WER on Common Voice test set. Handles 16kHz audio input.
Brief-details: PsychBERT is a BERT-based model specialized for mental health and psychology, trained on 40K PubMed papers and 200K social media conversations about mental health.
Brief-details: Dutch RoBERTa-based model fine-tuned for toxic comment detection, achieving 95.63% accuracy and 78.8% F1 score on Jigsaw dataset
Brief-details: BERT-based NER model specialized in identifying cities and countries in text. Fine-tuned on custom dataset with 3-tag classification system (OTHER, CITY, COUNTRY).
Brief-details: Large-scale Swin Transformer vision model pre-trained on ImageNet-21k, optimized for 384x384 images using hierarchical feature extraction and local attention windows.
Brief-details: MPNet-base is a pre-trained transformer model by Microsoft that excels at sentence embeddings and text similarity tasks, optimized for efficient processing.
BRIEF-DETAILS: Microsoft's MarkupLM-large: A multimodal pre-trained model combining text and markup language for document understanding and information extraction tasks
Brief-details: A specialized image generation model focused on producing highly realistic outputs, created by digiplay. Available on CivitAI and Hugging Face, optimized for photorealistic results.
Brief-details: Vision Transformer model trained with DINO self-supervised learning, featuring 85.8M params and 224x224 input size. Ideal for image classification and embeddings.
Brief-details: A quantized 32B parameter language model based on Qwen architecture, optimized through distillation and AWQ compression for improved efficiency while maintaining strong performance.
BRIEF-DETAILS: A compact model by katuni4ka hosted on HuggingFace, potentially designed for lightweight applications with granite-related processing or classification tasks.
Brief Details: KoELECTRA small discriminator model - Korean language model based on ELECTRA architecture, optimized for efficient NLP tasks with reduced parameters
Brief-details: BiRefNet-portrait is a specialized AI model for high-resolution portrait matting, achieving impressive performance metrics (0.983 Smeasure, 0.996 maxFm) on the TE-P3M-500-P dataset
