Checkpoint sweep comparison for tipsv2_token_long (ck20 to ck80) with arisole-eval
I ran ck20 vs ck60 from patch_policy_libero10_tipsv2_token_long_eval through arisole-eval. Both parse (n=100 each). Success is 33.0% vs 57.0%, delta +24.0 pp, SE 6.8 pp, unpaired because the files carry no per-episode seeds. These deltas are descriptive for a checkpoint sweep; checkpoints in one sweep are not independent runs, so I am not making a significance claim.
The sweep layout (ck20/40/60/80 plus nas5/8/12) is a good fit for a pairwise gate: it can flag a later checkpoint that is not better than an earlier one, with the uncertainty attached. Were all checkpoints evaluated on the same initial states / episode identity? If yes, a seed or episode-index field in eval_info.json would let the comparison be paired. A shared evaluation seed alone would not prove pairing.
To run it yourself: pip install arisole-eval==0.1.4, add run_meta.json per run directory, then arisole evaluate --baseline ck20_dir --candidate ck60_dir --out r.json && arisole report r.json. I am happy to adjust the tool if your layout needs it.
Disclosure: I work on arisole-eval (Apache-2.0, built by Arisole). It is not an official LeRobot or Hugging Face project.