# Synthetic Data Training

A vision-language model specialized in describing anime-style images, fine-tuned based on SmolVLM-500M-Base

A video caption generation model based on VILA-v1.5-8B, capable of generating detailed and human-preference-aligned captions for input videos.

PoseLess is an innovative robotic hand control framework that directly maps 2D images to joint angles using projection representations, eliminating the need for explicit pose estimation.

Poseless-3B is a vision-language model (VLM)-based robotic hand control framework that directly maps 2D images to joint angles without explicit pose estimation.

GLiNER-BioMed is an efficient open NER model suite based on the GLiNER framework, specifically designed for the biomedical domain to recognize various types of biomedical entities.

GLiNER-BioMed is an efficient open biomedical named entity recognition model suite based on the GLiNER framework, specifically designed for the biomedical domain, capable of recognizing multiple entity types.

GLiNER-BioMed is a specialized and efficient open biomedical NER model suite based on the GLiNER framework, achieving state-of-the-art zero-shot and few-shot performance in biomedical entity recognition tasks.

Asagi-8B is a large-scale Japanese Vision-Language Model (VLM) trained on extensive Japanese datasets, integrating diverse data sources.

A high-quality speech language model trained on a single GPU within 24 hours, fine-tuned based on Qwen2.5-0.5B, using Hubert tokens as vocabulary

A text classification model fine-tuned based on ModernBERT-large, designed to detect and classify various types of bias in text.

Asagi-14B is a large-scale Japanese Vision and Language Model (VLM) trained on a wide range of Japanese datasets, integrating diverse data sources.

A ControlNet model developed by Jasper Research Team for low-resolution image upscaling

A multilingual sentiment analysis model fine-tuned based on DistilBERT, supporting 21 languages, suitable for various scenarios such as social media and customer feedback analysis.

A multimodal large language model specifically trained to enhance low-level geometric perception, improving geometric analysis capabilities through high-fidelity synthetic visual descriptions

mStyleDistance is a multilingual style embedding model designed to closely embed texts with similar writing styles while distancing those with different styles, regardless of content or language constraints.

A text summarization model fine-tuned based on pegasus-x-base, trained with synthetic data, excelling in long document summarization tasks.

A ControlNet model developed by the Jasper research team for low-resolution image upscaling

Reflection Llama-3.1 70B is an open-source large language model trained with 'reflection tuning' technology, capable of autonomously detecting reasoning errors and correcting its approach.

A fine-tuned version of Depth Anything V2 for indoor metric depth estimation using the synthetic Hypersim dataset, compatible with the transformers library.

A version fine-tuned for indoor metric depth estimation tasks using the Hypersim synthetic dataset, based on the Depth Anything V2 model

A model fine-tuned from Depth Anything V2 for indoor metric depth estimation tasks, trained on the synthetic dataset Hypersim, compatible with the transformers library.

A fine-tuned version of Depth Anything V2, specifically designed for metric depth estimation in outdoor scenes, trained on the synthetic dataset Virtual KITTI.

A version fine-tuned for outdoor metric depth estimation tasks using the synthetic Virtual KITTI dataset, compatible with the transformers library.

A sentiment analysis model fine-tuned based on distilbert/distilbert-base-uncased, trained solely on synthetic data, supporting 5 sentiment classifications.

StyleDistance is a style embedding model designed to closely embed texts with similar writing styles and distance those with different styles, unaffected by content.

An 8.9 billion parameter language model developed in the third iteration using self-play preference optimization, starting from google/gemma-2-9b-it and fine-tuned with the UltraFeedback dataset

A specialized summarization model fine-tuned for 2 rounds based on Qwen2-1.5B-Instruct

TrOCR-Ru is an optical character recognition model fine-tuned on synthetic datasets of Russian and English, based on microsoft/trocr-base-handwritten, focusing on image-to-text tasks.

Merlinite 7B is a language model developed based on Mistral-7B-v0.1. It is trained using the LAB alignment method developed by IBM Research and performs excellently in multiple benchmark tests.

A zero-shot classification model based on the RoBERTa architecture, designed for text classification tasks without requiring training data, supports both GPU and CPU operation, and is trained using fully business-friendly data.

Zephyr 7B Gemma is a language model fine-tuned based on google/gemma-7b, trained on publicly available synthetic datasets using Direct Preference Optimization (DPO), designed to serve as a helpful assistant.

A Russian and English OCR model fine-tuned from microsoft/trocr-base-handwritten, specializing in handwritten and printed text recognition

The OpenMath model solves mathematical problems by combining textual reasoning with Python interpreter-executed code blocks, fine-tuned based on Mistral-7B-v0.1

ko-deplot is a Korean visual question answering model based on Google's Pix2Struct architecture, fine-tuned from the Deplot model, supporting chart image question-answering tasks in Korean and English.

Orca 2 is a research-oriented language model developed by Microsoft, focusing on enhancing the reasoning capabilities of small language models.

Orca 2 is a research-oriented language model developed by Microsoft, focusing on enhancing the reasoning capabilities of small language models, fine-tuned based on LLAMA-2.

This model is used to extract structured information from EU driver's license images and return the results in JSON format.

TrOCR is a Korean image-to-text model based on a vision encoder-decoder architecture, using DeiT as the image encoder and RoBERTa as the text decoder.

An instruction generation model fine-tuned on the deduplicated version of Pythia-2.8B, optimized for synthetic instruction datasets

OCTFusion is a PyTorch-based image classification model that achieved 100% accuracy on synthetic data.