How to Post-Train
A running series of tutorials, opinionated pieces, and guides around what actually goes wrong in post-training. Everything from trajectory eyeballing, rubrics and verifiers, task design, to RL environment quality. Written from years in the trenches :).
Model behavior notes
RL Fundamentals Mini-Series
-
01
LiveHow to Eye Ball Trajectories: You’ve Never Spent Real Time with Your Model and We Can ALL TellA no-BS guide for startups post-training their own models
-
02
LiveRL Environment Harness Quality: Stop Shipping Low-Quality Harnesses and Calling It an “Environment”Flaky harnesses quietly ruin your training runs; a few interactive examples
-
03
LiveRubric Quality pt1: Your Rubric Was Written by Someone Who Has Never Done the JobCheck that your evals are measuring the right thing
-
04
LivePM POV on Post Training: Your Tasks Are Not Grounded in Economic RealityGround your task design in what actually moves the needle for your product space
-
05
LiveData Quality pt 1: Your Data Screams “This Is a Simulation”Models detect fake environments and the behavior degrades fast
Reference
Who this is for
- Startups post-training their own models You have a training run, a harness, and a growing suspicion that something in your data is quietly wrong. These posts give some notes of what to look for.
- Data and eval teams On writing rubrics, critics, and state optimizers a practitioner would actually use, and building verifiers that measure the thing you actually care about.
- Anyone eyeballing trajectories How to read a trajectory, build a failure taxonomy, and tell a model failure apart from a harness failure before you burn a run on it.
- People buying RL data What to ask for, what to reject, and the tells that a vendor’s environment was built by someone who Googled the domain for an afternoon.