Page 3 of 8~104 min topic

Alignment and RLHF

Learn the controls and knobs

Each alignment / RLHF control is a hypothesis about a metric under a workload—not a synonym for quality on the general-purpose chat model.

~13 min this pageControls

1Learn the idea

Read

Control map

Primary knobs for alignment / RLHF: rater guidelines, preference mixture, KL / stay-close penalty, safety refusal policy, eval harm suite, over-refusal checks.

Write a sheet for the general-purpose chat model with columns: control, current value, predicted benefit, predicted cost, rollback trigger. Fill it using this topic’s real tension: Stronger safety can increase over-refusals. Heavier RL can reduce diversity or degrade niche skills. More rater hours improve signal and raise cost and cultural bias risk.

Change one alignment / RLHF family at a time. If you move two knobs and the general-purpose chat model improves, you learned a cocktail, not a cause—and you cannot roll back surgically.

Read

Product exposure

End users of the general-purpose chat model should see only safe dials related to alignment / RLHF. Infrastructure limits, private prompts, and policy thresholds stay server-owned. A user-facing control that bypasses those limits is a vulnerability dressed as UX for alignment / RLHF.

Read

Make it operational

Publish the alignment / RLHF control sheet next to the general-purpose chat model runbook. On-call should see which knob moved in the last deploy without reading chat archaeology. Unknown alignment / RLHF knobs are unowned knobs.

Also pin one numeric memory from this alignment / RLHF chapter: If raters prefer answer A over B in 62 of 100 pairs, the preference model should rank A higher on held-out pairs; a 50/50 split means the signal is noise for that slice. That number is not decoration; it is a template for how claims about alignment / RLHF on the general-purpose chat model should look in design docs. Scoped specifically to alignment / RLHF / general-purpose chat model / controls-and-knobs.

Read

Common mix-ups

People confuse alignment / RLHF with neighboring buzzwords when debugging the general-purpose chat model. Before changing prompts, ask whether the broken stage was evidence gathering, the alignment / RLHF judgment itself, validation, or the product action. Fixing the wrong stage creates folklore (“we tried alignment / RLHF and it failed”) that blocks the next team on the general-purpose chat model. Scoped specifically to alignment / RLHF / general-purpose chat model / controls-and-knobs.

Read

Rehearsal (alignment-and-rlhf/controls-and-knobs)

Write a five-line artifact for this page: goal, inputs, check, owner, stop rule. Invent one fluent failure that the check would catch. Keep details specific to alignment and rlhf rather than generic AI advice.

Read

Rehearsal (alignment-and-rlhf/controls-and-knobs)

Write a five-line artifact for this page: goal, inputs, check, owner, stop rule. Invent one fluent failure that the check would catch. Keep details specific to alignment and rlhf rather than generic AI advice.

Go deeper

Before you start

Why this matters

From [rater guidelines, preference mixture, KL / stay-close penalty, safety refusal policy, eval harm suite, over-refusal checks], pick one control for alignment / RLHF on the general-purpose chat model. Predict which metric rises and which cost rises if you increase it.

Check your understanding

Page assessment

Answer from memory. Completion is saved from this evidence, not from opening the next page.

1. What is one idea from this page you would apply, and what evidence would you check?

All responses are required.