RoboTraining-raw / README.md
Sssunset's picture
Raw scene records of Odd Parcel (text version) for the trainer
85cdfcf verified
|
Raw History Blame Contribute Delete
1.4 kB
---
pretty_name: RoboTraining raw records
---
# RoboTraining raw records
This dataset has the raw scene records that the Posterior Learning trainer reads for the text version of Odd Parcel
(`evidence/parcels.py` in the pro-worlds repository). The scenes are development scenes from the benchmark's own generator.
They are the same scenes as in `Sssunset/RoboTraining`. There are no test scenes. The test is the benchmark's eval50
instances in the simulator.
## Files
- `snap/snap-0.jsonl` to `snap/snap-23.jsonl` have 630 lines. Each line is one scene that the simulator built once. The line
records the goal, the intro and the action list of the benchmark's adapter, the parcels, the hidden parcels, the mass of
each parcel and the real balance readings.
- The demonstration list repeats 48 registry scenes. Thus the files have 582 different scenes. The trainer removes one scene
(`0e5cfae48cf5295b`) because the simulated balance sticks at its stop. 581 scenes remain.
## Splits
The trainer makes the splits from the registry file in the repository
(`benchmarks/roboquest/suite/v1/odd_parcel/registry/registry.json`):
- Validation: every fourth development row of the registry, 33 scenes.
- Training: all the other scenes, 548 scenes.
## Use
```bash
hf download Sssunset/RoboTraining-raw --repo-type dataset --local-dir parcels-scenes
export PARCELS_SNAP=$PWD/parcels-scenes/snap
```