I silently break training codes or configs so I made pybench [P]

It is like pytest but for statistical tests: it ensures no regression of your metrics at a statistical level. It manages tedious things such that seeds, past benchmark results, ... Simple CLI working like pytest but with benchmarks/ directory instead of tests/: pybench # 1st time: samples seeds, saves a baseline, marks NEW pybench # later: reruns on the same seeds, marks PASS / FAIL pybench update # re-baseline after an intended change pybench show # print current baseline stats (--history for per commit) P