Overview

DatasetLoom is an innovative platform designed specifically for AI engineers, researchers, and teams looking to create high-quality multimodal training datasets. This intelligent platform streamlines the entire process from raw data to structured training samples, supporting various tasks such as supervised fine-tuning (SFT), preference alignment (DPO), image captioning, and visual question answering (VQA). Its modular design and visual interface enhance efficiency and make it easier for users to build and evaluate datasets with diverse data types.

Whether you’re working with text, images, or both, DatasetLoom provides the tools necessary to facilitate the generation and evaluation of complex datasets. It empowers teams to compare model outputs and automatically assess their quality, ensuring comprehensive support for multimodal models.

Features

• Multimodal Dataset Construction: Generates training data encompassing images, text, and visual question answering, essential for diverse AI model training tasks.

• Model Evaluation and Scoring: Offers AI-powered automatic scoring, multi-model comparisons, and comprehensive quality assessments to gauge model performance effectively.

• Document Parsing: Supports uploads and extraction from PDF, Word, Markdown, and TXT formats, enabling seamless integration of various document types into the workflow.

• Image Annotation and Chunking: Features tools for image region labeling and generation of text-image descriptions, essential for creating robust visual datasets.

• User and Permission Management: Simplifies user management with login, registration, and role assignments, ensuring controlled collaboration among team members.

• Data Persistence: Automatically saves dialogue history, question generation, and dataset versions, facilitating easy access to past work and dataset iterations.

• Training Data Export: Supports exporting datasets in JSON, CSV, and HuggingFace Dataset formats, providing flexibility for further model training and integration.

• Workflow Engine (Beta): Introduces a Redis-based asynchronous task scheduling system that automates complex processes, optimizing pipeline efficiency.