All learning paths Learn See Try Make Where your labels go What this lesson is about Follow a rating batch from your screen to a training run, and find the checks it passes on the way.
You finish two hundred rows on Friday and press send. Your cousin asks whether a model changes that evening.
A Each row reaches the model as it is sent, so the change starts the same night. B Your batch is pooled with others and then split: most goes to the run, some is kept aside. C The rows sit in a store until somebody asks for them, and nothing is set apart.
Check my recall Why this matters A published assistant run describes the loop plainly. Batches are pooled, a part is kept aside, the model is trained, and fresh batches arrive on the changed model. Two things decide whether the run can be believed later. A held-out set — items the model never trains on — must stay unread during repairs. And every batch must carry who marked it and when.
Follow one rating batch from your screen to a training run, and name the checks on the way.
A published assistant run describes the loop plainly. Batches are pooled, a part is kept aside, the model is trained, and fresh batches arrive on the changed model. Two things decide whether the run can be believed later. A held-out set — items the model never trains on — must stay unread during repairs. And every batch must carry who marked it and when.
Your batch is pooled with others and then split: most goes to the run, some is kept aside. The part kept aside is spent the moment somebody reads its failures to make a repair. Write who marked it and when, because a batch with no name cannot be checked later. Task A team asks where last month's twelve hundred marked pairs went before the new model was tested.
Weak approach Says they all went into the training run, and the new model answers better, so the marking worked. Names no batch, no count and no date.
Stronger approach Traces them: a thousand into the run, two hundred kept aside. Two of the two hundred were read during repairs, so the report says one hundred and ninety-eight.
Why the stronger approach works The weak trace stops at the model feeling better. The strong one says what was kept aside and what was spent.
Try a changed situation A partner wants one number from the part you set apart. Six of its items were read while repairing the guide.
Make something yourself Draw the path of one batch you worked on. Who marked it, what was trained on, what was kept aside, what was spent.
Lesson 8 of 9 on How post-training works next Audit a preference batch Part 2 ends in a work sample you can send
Read the primary or official source: Bai et al., Training a Helpful and Harmless Assistant with RLHF (arXiv)