How-to guides
You know the job and you want the exact calls. Each recipe is one page, one task, and code you can paste. The runnable ones use the tiny CPU signal, so they finish in under a minute with no download; anything that needs an 8B model is marked and names the call rather than pretending to run.
If you want to understand why the calls are shaped the way they are, the instruments and concepts are the longer read.
Point reward-lens at your grader
-
Wrap a model already in memory, or build a tiny one on CPU. The hub loader is gated and says so.
-
Use a DPO checkpoint as a signal
A DPO policy is a reward model with no head. Read its implicit log-ratio reward.
-
Read a verdict off the unembedding, and validate that the judgment lands where you think.
-
Score with a process reward model
One reward per reasoning step. The delimiter split is exact; the learned detector is a stub.
-
Make a model family reward-lens does not ship support for speak the signal protocol.
Read an instrument
-
Run the bias battery across axes and read a standardized effect size with an honest sample size.
-
Split a margin into per-component contributions, and know why that is only the first look.
-
Patch without running out of memory
Triage with the cheap tools, then patch only the components that matter.
-
See where two models form the same preference, and why raw numbers need a shared frame first.
Trust the number
-
Effective sample size of an eval set
Thirty rows from six seeds are not thirty data points. Count what the data actually earned.
-
Calibrate a detector on an organism
No instrument earns more than exploratory trust without an answer key to check it against.
-
Write the prediction down first, freeze it, then let the spec adjudicate the result.
-
Build a card, check a manuscript
A card can only say what the evidence store can back. A claim check fails on an unbacked number.
Wire a training loop
-
Attach a reward function and log the geometry as optimization pressure builds.