Alignment and RLHF
Weigh the tradeoffs
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.
1Learn the idea
Read
The live 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.
Translate into user impact on the general-purpose chat model when tuning alignment / RLHF. Which error class costs more—missed catches, slower answers, higher spend, or privacy exposure? That ranking picks the default more honestly than a blog’s recommended settings for alignment / RLHF.
Read
Numbers that force honesty
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. Scoped specifically to alignment / RLHF / general-purpose chat model / tradeoffs.
If the aggressive alignment / RLHF setting wins the headline metric while breaking a protected slice or blowing the latency budget on the general-purpose chat model, it is not a win. Record intended gain and tolerated regression together for alignment / RLHF.
Read
Make it operational
Revisit the alignment / RLHF tradeoff when traffic shape changes on the general-purpose chat model. A setting that was right at low volume can fail when a new language segment or document length appears. Tradeoffs expire; re-measure on a calendar, not only on incidents.
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 / tradeoffs.
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 / tradeoffs.
Read
Rehearsal (alignment-and-rlhf/tradeoffs)
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
For the general-purpose chat model, name one regression you will tolerate when pursuing the main benefit of alignment / RLHF, and one regression that is stop-ship.
Related lessons
Check your understanding
Page assessment
Answer from memory. Completion is saved from this evidence, not from opening the next page.
All responses are required.