Models

BRIEF DETAILS: A TinyBERT model fine-tuned on MS MARCO dataset, optimized for efficient passage ranking and information retrieval tasks.

Brief-details: English to Tagalog (Filipino) neural machine translation model based on transformer architecture, achieving 26.6 BLEU score on Tatoeba test set

BRIEF-DETAILS: Quantized versions of stable-code-instruct-3b optimized for different performance/size tradeoffs, ranging from 1.08GB to 2.97GB with varying quality levels

Brief Details: GLaMM-GranD-Pretrained: Advanced multimodal model trained on 7.5M concepts across 810M regions for detailed visual understanding and segmentation.

BRIEF DETAILS: Specialized Chinese text emotion classifier based on mDeBERTa-v3, capable of identifying 8 distinct emotional tones. Lightweight version (small) optimized for efficient deployment.

Brief-details: Japanese sentence embedding model based on LUKE architecture, optimized for semantic similarity tasks with comparable or better performance than BERT-based alternatives

Brief-details: A PEFT-optimized tiny OPT model using LoRA adaptation, designed for testing and development purposes. Built with PEFT 0.4.0.dev0 framework.

Brief-details: A specialized token classifier model developed by Danswer, combining intent classification and token identification capabilities for enhanced natural language understanding tasks.

Brief-details: MedEmbed-large-v0.1 is a specialized embedding model fine-tuned for medical and clinical information retrieval, offering superior performance on healthcare NLP tasks compared to general-purpose models.

Brief-details: A fine-tuned wav2vec2-xls-r-300m model for Finnish ASR, trained on 275.6 hours of Finnish speech data achieving 17.92% WER without LM and 8.16% with LM.

Brief-details: ESPnet speaker recognition model trained on VoxCeleb, achieving 0.739% EER. Uses RawNet3 architecture with self-supervised front-ends for speaker embeddings.

Brief Details: State-of-the-art text-to-video model with 30B parameters. Features 16x16 spatial and 8x temporal compression, generating up to 204-frame videos with DPO optimization.

Brief-details: Highly accurate English ASR model trained on 200K hours of human-transcribed speech, featuring adjustable verbatimicity and multiple decoding options

Brief-details: A powerful Mistral-based 7B parameter model optimized through SLERP merging of leading models, designed specifically for downstream fine-tuning tasks and superior performance.

Brief-details: T5-large-lm-adapt is Google's enhanced T5 model with GEGLU activation, improved pre-training on C4, and specialized LM adaptation for better prompt tuning capabilities.

Brief-details: T5-efficient-tiny-nl32 is a deep-narrow variant of T5 with 67M parameters, optimized for efficiency through increased depth (32 layers) while maintaining a narrow architecture

Brief Details: T5-efficient-small is a deep-narrow variant of T5 with 60.52M parameters, optimized for efficiency through increased depth rather than width.

Brief Details: T5-efficient-base is a deep-narrow variant of T5 with 223M parameters, optimized for efficient scaling and downstream performance through increased model depth.

Brief-details: RemBERT is Google's multilingual BERT variant trained on 110 languages with separate input/output embeddings, optimized for classification tasks

Brief Details: BERT-based multilingual model pre-trained on 17 Indian languages, optimized for both native and transliterated text processing, achieving strong cross-lingual performance.

Brief Details: FNet-large is Google's transformer variant using Fourier transforms instead of attention, trained on C4 with 24 layers and 1024 hidden dimensions for MLM/NSP tasks.