Training Evaluation: Fill & Download for Free

Green Team is a training/evaluation unit within DEVGRU (formerly known as SEAL Team Six). The SEALS in Green Team are not yet members of DEVGRU. In essence, they are trying out for membership in this elite, Tier One unit.DEVGRU candidates remain in this intensive program, constantly under the watchful eye of instructors, for at least six months.The typical attrition rate is approximately 50%.If selected, the successful SEAL candidate will be placed in one of the handful of units that comprise DEVGRU.** Members of the DEVGRU Red Squadron on a training mission.

Here are some rules of thumb.Cross-validation is almost always preferable to a straight-up train/evaluate/test split.3 to 5 random folds is typically sufficient. Sometimes 10 can be better if you have the patience for it.Reserve a final fold for testing. This is typically larger than one of the cross-validation folds but smaller than them combined.If you have time series data, i.e. data where the time order matters, you need to do forward chaining. That is, you train on fold 1, evaluate on fold 2, then train on fold 1+2, evaluate on fold 3 and so on. Random splits is a big no-no with time series data.If you do forward chaining, you need to be very careful that you avoid any data leakage between the folds. Even if all of the samples in fold 1 were collected before the samples in fold 2, it could be the case that some samples in fold 1 share information implicitly with samples in fold 2. This type of data leakage can be extremely subtle.The same type of caution with regards to data leakage should be used when splitting the cross-validation set from the test set.Always use evaluation score or cross-validation score when comparing models. Never compare based on their score on the training set.Finally, and this is something that almost everyone seems to miss:The real test is to run the model in production and observe how well it performs on live data. Model performance almost always declines like this: Training score > Cross-validation score > Test score > Live data score. Real-world data distributions change over time, and this is just what happens.

I can help you out. I usually work with companies, but I can also take on individuals. My specialty is software development and product management interviews.I offer two types of consulting services:Helping companies prepare for their acquisition/acqui-hire interviews.Interviewing training/evaluation: I "shadow" interviews and help interviewers design good technical questions.I'm a former Google/Microsoft/Apple software developer and the author of Cracking the Coding Interview.From linkedin.com