Clarifying the split column vs. Alpamayo-R1 training data. Which clips are safe as a held-out eval set?
#36
by kimseungho - opened
Hi, and thanks for releasing this dataset! We're curating a subset to evaluate Alpamayo-R1 checkpoints, so avoiding any train/eval contamination is important to us. A few clarifications would help:
- Dataset-wide split semantics.
clip_index.parquetassigns a split to all 306,152 clips (we see ≈ train 153,625 / val 90,928 / test 61,599). Is this the official train/val/test partition for the full dataset, and is it intended for general model development? The dataset card only documents splits for the OOD reasoning subset, so we want to confirm the dataset-wide column's meaning. - Relationship to Alpamayo-R1 training. The Alpamayo-R1 report states the model was trained on an internal ~80,000-hour corpus, with the public PhysicalAI-AV set used for evaluation (Table 10). Given the shared collection program (same cities/rig), can you confirm whether any clips in the public train (or val/test) split were included in — or overlap with — the training data of the released AR1 checkpoints? In other words, which split(s) are guaranteed disjoint from AR1's training set and therefore safe to use as a held-out test set?
- Identifying the official evaluation subsets. The AR1 report (Table 10) reports results on "644 examples from the PhysicalAI-AV evaluation set" (open-loop) and "920 scenarios from the PhysicalAI-AV NuRec dataset" (closed-loop / AlpaSim). Is there a manifest, flag, or filter in the dataset that identifies these specific clips, so we can reproduce/extend that exact evaluation set?
- NuRec subset. The README notes NuRec-based reconstruction is supported only for the Hyperion 8.1 sensor rig. How can we filter the dataset to the NuRec-eligible clips/scenarios used for closed-loop evaluation?
Thanks in advance!
kimseungho changed discussion title from Clarifying the split column vs. Alpamayo-R1 training data — which clips are safe as a held-out eval set? to Clarifying the split column vs. Alpamayo-R1 training data. Which clips are safe as a held-out eval set?