metadata
pretty_name: Spa-Bench Data Information
task_categories:
- robotics
tags:
- spa-bench
- evaluation
- metadata
configs:
- config_name: model_evaluation
data_files:
- split: train
path: Model Evaluation Data Information.csv
- config_name: training_data
data_files:
- split: train
path: Training Data Information.csv
Spa-Bench data information
This repository exposes two reviewer-friendly tables through Hugging Face Data Studio:
- Model Evaluation Data Information: 2,892 scored physical rollouts across five checkpoints.
- Training Data Information: the 1,200-row prompt and compositional-design manifest.
Use the configuration selector in the viewer to switch between the two tables. Task names are normalized to Counting, Ordinal Position, Physical State, Referential Description, Relational Placement, and Relative Size.
Training-design summary
All atomic concepts and object identities appear during training. The benchmark withholds selected concept–argument combinations rather than unseen words, objects, or motor outputs. Each task family contains 200 demonstrations.
Models
- GR00T-N1.7 Frozen LLM — Spa-Bench Epoch 12
- GR00T-N1.7 Full Fine-Tune — Spa-Bench Epoch 12
- VLA-0 — Spa-Bench Epoch 12
- MolmoAct2 — Spa-Bench Step 76,596
- pi0.5 — Spa-Bench Step 76,596
Evaluation rollouts
- Spa-Bench Evaluation Rollouts — pi0.5 (Full)
- Spa-Bench Evaluation Rollouts — MolmoAct2 (Full)
- Spa-Bench Evaluation Rollouts — GR00T-N1.7 Frozen LLM (Full)
- Spa-Bench Evaluation Rollouts — GR00T-N1.7 Full Fine-Tune (Partial)
- Spa-Bench Evaluation Rollouts — VLA-0 (Partial)
Training dataset
Double-blind release note
Author, institution, source-repository, and archival citation details are intentionally omitted during review. They will be restored in the archival release.