SN80 · OpenRoboto

checkpoint merge toolkit
loading…

Score of every submission

The leaderboard shows only the champion and dethroned ex-champions — a failed challenge never appears on it. These are all submissions, recovered from the public queue.

What it takes to earn anything

Merge pool

Queue

Base model what every submission was fine-tuned from

Submissions

Every miner LoRA-tuned the same π0.5 base, so a submission's delta from that base is its merged adapter. What decides whether two checkpoints can be combined is not the score — it is the parameter tree, shown here for each.

uidMiner / repoScorevs base ArtifactParameter treeMerge verdict

Merge planner

Pick the checkpoints to combine, set the recipe, and get the exact command plus a check of the three ways a merge here fails silently.

UseModelScoreWeightStatus

Recipe

Commands

What the total score is actually made of

The production profile libero_pro_custom_1 reports six entries and averages them equally — but it runs sixteen. The four base groups each average their four perturbation dimensions, then libero_object_swap and libero_spatial_swap are appended a second time as standalone entries. Those two therefore carry 5/24 of the total each.

Where the headroom is

Published π0.5 success rate per perturbation dimension. The base behaviours are saturated; the swap dimension is both the weakest and the most heavily weighted.

Margin calculator

How much you must move the two double-weighted swap suites to clear the champion by 0.02.

Score calculator enter per-suite success rates, get the reported total

Seeded with the published π0.5 numbers. Change any suite to see what it buys, or paste your own run_eval.py results.

SuiteWeightSuccess

Submission rules that cost money to get wrong

Pre-flight checklist

Toolkit reference README.md from /root/sn80-merge

loading…