RE/RS, Data Understanding (MM)

The Data Understanding team is responsible for creating the high quality datasets and their quantized representation for OpenAI. This includes synthesizing multimodal data, building VQ representations, and processing, filtering, deduplication, quality control, and tokenization so it can be used effectively in big model training runs. We’re looking to advance how OpenAI prepares, curates, synthesizes and understands multimodal data at scale. You’ll work on research and production problems like synthesizing multimodal content (images, audio, and video) and their supervisions, improving noisy data pipelines, building better quality filters, using models to automate data prep, and measuring whether changes in the dataset improve model performance.

Responsibilities

Advance how OpenAI prepares, curates, synthesizes and understands multimodal data at scale; Work on research and production problems like synthesizing multimodal content (images, audio, and video) and their supervisions; Improve noisy data pipelines; Build better quality filters; Use models to automate data prep; Measure whether changes in the dataset improve model performance.

Requirements

Have a strong track record of new or improved ML ideas, through publications, projects, or applied research; Own and drive a research agenda, from choosing the right multimodal data problems to carrying long-running work through to impact; Be excited by OpenAI’s empirical, collaborative approach to research.

Nice to have

Experience with multimodal learning, audio, vision, video, synthetic data, or data-centric ML; Thoughtfulness about AI’s impact, including privacy, provenance, and data quality; Experience building high-performance deep learning or large-scale data processing systems.