RecipeBook by Shofo

Recipe Book is a comprehensive video data platform designed for individuals and teams who need to acquire high-quality video training data for their AI models. It offers a streamlined process for searching, filtering, and purchasing datasets, enabling users to obtain precisely what they need without lengthy procurement cycles. The traditional methods of acquiring training data, whether building it in-house or buying from vendors, are often time-consuming and resource-intensive. Building data requires significant engineering effort for collection, filtering, and maintenance, while buying data can involve high costs, sales calls, and extended iteration periods. Recipe Book addresses these pain points by providing a direct, efficient, and cost-effective solution for obtaining video training data. One of the core features of Recipe Book is its semantic search capability, which allows users to find relevant video clips using natural language queries. This goes beyond simple keyword matching, enabling users to discover content based on the meaning and context within the videos. Coupled with advanced metadata filters such as frames per second (fps), resolution, duration, and aspect ratio, users can precisely narrow down the vast catalog to find clips that meet specific technical requirements. Recipe Book also offers a unique iterative feedback loop. After an initial search, users can rate a few selected clips by upvoting or downvoting them based on their specific needs. This feedback is used to train a custom probe that re-ranks the entire catalog, ensuring that the results are tailored to the user's exact preferences and requirements. This iterative process allows for fine-tuning the dataset to a high degree of accuracy and relevance. The platform simplifies the purchasing process significantly. Once users are satisfied with the search results, they can specify the desired number of hours of video data. Recipe Book then provides a review sheet that includes the score distribution, a grid of the lowest-scoring clips for quality assessment, and a trim slider to finalize the selection. This transparency ensures users know exactly what they are buying. Acquiring the data is designed to be fast and straightforward. After confirming the purchase, users receive a manifest CSV file containing download links for the selected video clips. This process can be near-instantaneous for smaller purchases, eliminating the need for contact forms or sales calls, and offering a clear price of $3 per hour. Recipe Book's methodology is built on years of experience in data collection and curation, indexing billions of videos across the internet. The platform leverages this extensive indexing to make publicly available data accessible and usable for AI model training. The approach focuses on visual search, as metadata like captions and engagement data are not yet comprehensively integrated into the corpus. The primary benefit for users is the ability to acquire custom video training datasets rapidly, often within the same day. This accelerates the model development lifecycle, allowing teams to move from data sourcing to training and iteration much faster, potentially achieving state-of-the-art results more quickly. Specific use cases include training video generation models, Vision-Language Models (VLMs), and world models. For instance, a user might search for specific cooking motions to train a robot chef, or for particular environmental scenes to enhance a self-driving car's perception system. The platform is ideal for solo developers, small teams, or larger organizations looking to quickly source data for AI projects. Recipe Book is available as a web platform. The pricing is set at $3 per hour for self-serve purchases, with options for custom contracts and indemnity discussions for larger deals. The target audience includes teams training video or multimodal models, particularly those in areas like video generation, VLMs, and world models, from seed-stage startups to Series A companies. In summary, Recipe Book by Shofo revolutionizes video data acquisition by offering a fast, semantically searchable, and cost-effective platform for obtaining precisely tailored training datasets, empowering AI development with unprecedented speed and efficiency.