The electroencephalogram (EEG), which has been in clinical use for over 70 years, is still an essential tool for diagnosis of neural functioning (Kennett, 2012). Well-known applications of EEGs include identification of epilepsy and epileptic seizures, anoxic and hypoxic damage to the brain, and identification of neural disorders such as hemorrhagic stroke, ischemia and toxic metabolic encephalopathy (Drury, 1988). More recently there has been interest in diagnosing Alzheimer's (Tsolaki et al., 2014), head trauma (Rapp et al., 2015), and sleep disorders (Younes, 2017). Many of these clinical applications now involve the collection of large amounts of data (e.g., 72-h continuous EEG recordings), which makes manual interpretation challenging. Similarly, the increased use of EEGs in critical care has created a significant demand for high-performance automatic interpretation software (e.g., real-time seizure detection). A critical obstacle in the development of machine learning (ML) technology for these applications is the lack of big data resources to support training of complex deep learning systems. One of the most popular transcribed seizure databases available to the research community, the CHB-MIT Corpus (Goldberger et al., 2000), only consists of 23 subjects. Though high performance has been achieved on this corpus (Shoeb and Guttag, 2010), these results have not been representative of clinical performance (Golmohammadi et al., 2018). Therefore, we introduce the TUH EEG Seizure Corpus (TUSZ), which is the largest open source corpus of its type and represents an accurate characterization of clinical conditions. Since seizures occur only a small fraction of the time in this type of data, and manual annotation of such low-yield data would be prohibitively expensive and unproductive, we developed a triage process for locating seizure recordings. We automatically selected data from the much larger TUH EEG Corpus (Obeid and Picone, 2016) that met certain selection criteria. Three approaches were used to identify files with a high probability that a seizure event occurred: (1) keyword search of EEG reports for sessions that were likely to contain seizures (e.g., reports containing phrases such as “seizure begins with” and “evolution”), (2) automatic detection of seizure events using commercially available software (Persyst Development Corporation., 2017), and (3) automatic detection using an experimental deep learning system (Golmohammadi et al., 2018). Data for which approaches (2) and (3) were in agreement were given highest priority. Accurate annotation of an EEG requires extensive training. For this reason, manual annotation of EEGs is usually done by board-certified neurologists with many years of post-medical school training. Consequently, it is difficult to transcribe large amounts of data because such expertise is in short supply and is most often focused on clinical practice. Previous attempts to employ panels of experts or use crowdsourcing strategies were not productive (Obeid et al., 2017). However, we have demonstrated that a viable alternative is to use a team of highly trained undergraduates at the Neural Engineering Data Consortium (NEDC) at Temple University. These students have been trained to transcribe data for seizure events (e.g., start/stop times; seizure type) at accuracy levels that rival expert neurologists at a fraction of the cost (Obeid et al., 2017; Shah et al. in review). In order to validate the team's work, a portion of their annotations were compared to those of expert neurologists and shown to have a high inter-rater agreement. In this paper, we describe the techniques used to develop TUSZ, evaluate their effectiveness, and present some descriptive statistics on the resulting corpus.