Technical case studies and post-training experiments behind Nablo's workload-to-model product. The SQL series shares one frozen benchmark, so its numbers compare to each other. The methods that did not work are here too.
Post-training a 9B open model on one bounded workload until it reached the same score as GPT-5.5 medium, on a benchmark frozen before the work started.
No larger teacher at all. The student samples its own actions and trains on probability feedback from a frozen copy of itself.
A 35B teacher scores the student's own sampled actions with log probabilities instead of writing targets for it.
Teacher traces start from the task, so they never cover the mess the student's own policy creates. This one corrects the student at its own failure states, which is the method the 9B case study later scaled up.
Same student, same harness, one change. The student sees the top-k distribution behind each teacher choice instead of only the final pick.
A 0.8B model copies successful teacher trajectories one token at a time. It went from 1 solved task to 46, which taught it to operate the agent loop but not to write better SQL.