Dataset · Hugging Face

perpetual-research/splatbench

Name: splatbench
Creator: perpetual-research
License: other
Keywords: task_categories:reinforcement-learning, task_categories:image-to-3d, language:en, license:other, size_categories:10K<n<100K, format:parquet, modality:image, modality:tabular, modality:text, modality:3d, library:datasets, library:dask, library:polars, library:mlcroissant, region:us

SplatBench SplatBench is a benchmark and RL environment for raw spatial reasoning. A language model gets eight photos of an object and rebuilds the object in 3D by placing every Gaussian itself through tool calls, with no code and no optimizer. We grade the build from eight camera angles the model never saw, so only a correct 3D structure scores well. This repository holds the environment, the 30 tasks and all 1,230 episodes from the SplatBench report: every tool call, every… See the full description on the dataset page: https://huggingface.co/datasets/perpetual-research/splatbench.

View on Hugging FaceReport an issue

Author / organization	perpetual-research
License (as declared)	other
Downloads	208 (as reported by Hugging Face)
Likes	1
Created	13 October 2026
Last updated at source	13 October 2026

benchmarkrl-environmentspatial-reasoning3dgaussian-splattingagentstool-usetrajectories

Embed this entry

<iframe src="https://silytics.si/embed/item/77bc9076-e77a-42ce-9fc6-726e9e94dfbc" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Free, no tracking. More about feeds and embeds.

Popularity, last 3 days

208 now (+41 since 2026-10-16). Counts as reported by the source, snapshotted daily by Silytics.

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.