Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
duckbatch records
The records behind craigm26/duckbatch: batched, judged
distillation of Pollen Robotics' Microduck walker, plus decision-model and preference-learning
experiments. Everything was measured in simulation (mjlab / MuJoCo-Warp); nothing has run on a
robot. Each experiment was pre-registered in notes/ in the GitHub repo before it ran.
| Folder | What | Real or simulated |
|---|---|---|
records/b001-student-size |
Batch 1: every student killed at rung 0 (the early gates were too strict), with the rung eval, judge decisions and Jev/Decide shadow answers | real sim rollouts |
records/b002-student-size-longer |
Batch 2: two finalists (57,934 and 26,254 params), final held-out eval, bench, judge records | real sim rollouts |
records/r001-router-requests |
Plain-language router: GLiNER2.5-Decide and Jev on 46 pre-registered requests | real model outputs; gold labels written by hand |
records/p001-pairs |
Paired rollouts: 4 policies × 5 fixed commands × 512 envs; features and joint trajectories | real sim rollouts |
records/f000-simulated |
duck-feedback/0 example records |
SIMULATED: no person was involved |
Formats: duck-feedback/0 is documented in docs/duck-feedback-0.md, and the batch records
(duckbatch.batch.v1) in src/duckbatch/batch/runner.py, both in the GitHub repo.
- Downloads last month
- 20