Reticule

Curlset is a workspace for creating, labeling, reviewing, and exporting AI training datasets. We built it for developers, students, and small teams who need useful training data without building their own data tools. Asking an AI model to generate examples is easy. Preparing those examples for training is harder. They can be repetitive, miss important cases, or contain inconsistent labels. Teams still need to check the data, correct mistakes, track changes, and put everything into the right format. Curlset brings that work into one place. Users describe their task in plain language or upload existing data. They can generate text and image datasets, get AI-assisted labels, review individual records, and preview edits before applying them. A guided workflow helps beginners get started, while advanced controls let experienced users configure models and labeling settings. Our approach combines task-specific generation, diversity checks, multiple-model label checking, and human review. Quality checks flag duplicates, class imbalance, and overlap between training and test data. Every annotation keeps its source and review status, so users can distinguish model suggestions from human decisions. For image annotation, Meta SAM 3.1 proposes bounding boxes that users can inspect and correct. Once the data is ready, users can freeze a version and export it with a manifest describing its contents. Supported outputs include CSV, JSONL, image bundles, and formats for tools such as Unsloth. The current MVP focuses on text and images. Image quality and labeling accuracy still need further testing and improvement. We are also researching how small models trained on reviewed labels can help catch mistakes shared by general models. Our goal is to make reliable dataset preparation accessible without hiding the decisions that affect quality. Audio, video, tabular data, and other dataset types are part of our longer-term roadmap.