RSI-Dataset

Data splits of tasks of Flex-RSI, a benchmark that evaluates whether an agent, given a task and a fixed budget, can improve not only its solution to the task but also the harness through which it operates. The task packages (instructions, evaluators, starter solutions) and the benchmark itself are in the code repository, flex-rsi/rsi-bench; this repository holds the data the tasks' prepare.py produce.

Layout

<task>/
  dev/    available to the agent, with its labels
  val/    withheld from the agent: decides which candidate solutions are retained
  test/   withheld from the agent: scores the final solution once, after the run

The layout is the benchmark's data root. Download one task (or all) into it:

export RSIB_DATA_ROOT=/path/to/rsib-data
huggingface-cli download Flex-RSI/RSI-Dataset --repo-type dataset --include '<task>/*' --local-dir "$RSIB_DATA_ROOT"

Tasks

Task	--task	Source	Metric	Splits (dev / val / test)
3D Tracking	iggt-tracking	IGGT (ScanNet, ScanNet++)	T-mIoU	100 / 8 / 10 clips
Active Search	hstar-search	H*Bench	SPL	80 / 100 / 47 episodes
Visual Search	vsearch	V*Bench, HR-Bench 8K	Accuracy	70 / 80 / 140 questions
Video Quality	video-quality	YouTube-UGC, Veo 3.1 pairs	(SRCC + pairwise acc.) / 2	45 / 45 / 60 items
Image Editing	edit-image-unlimited	Ours	VLM judge + checks	50 / 70 / 80 images
Ego-Exo Memory	egoexomem	EgoExoMem (Ego-Exo4D)	Accuracy	80 / 80 / 120 questions (questions only)

The task cards (purpose, construction of the splits, setting, starter solution, metric) are in docs/tasks.md of the code repository.

Files

Each task's instruction and evaluator in the code repository define its files in full.

• iggt-tracking: <split>/inputs/<clip>/frames/*.png (RGB frames, 448 x 592, 6 to 8 per clip) and <split>/labels/<clip>/gt.npz (inst_ids: (S, 448, 592) int32, 0 = background, one id per object track); dev/manifest.json lists the scene and source frames of every clip.
• hstar-search: <split>/episodes.jsonl, panoramas/ and labels.jsonl; manifest.json lists the sources.
• vsearch: <split>/questions.jsonl, images/ and labels.jsonl; dev also has targets.jsonl.
• video-quality: <split>/items.jsonl (score one video, or compare two), videos/, frames/ and labels.jsonl (MOS, or the human-preference vote).
• edit-image-unlimited: <split>/samples.jsonl (instruction, edit family, checks), images/ (inputs, at most 1536 x 1024 px) and masks/.
• egoexomem: <split>/questions.jsonl, labels.jsonl and videos.jsonl; dev also has evidence.jsonl. The videos and frames are not included: build the full splits from your own copy of Ego-Exo4D with EGOEXO4D_ROOT=/path/to/egoexo4d rsib task prepare egoexomem.

License and terms of use

• iggt-tracking is derived from ScanNet (test) and ScanNet++ (val, dev); use is subject to the ScanNet and ScanNet++ terms of use.
• hstar-search: panoramas and episodes from H*Bench; its terms apply.
• vsearch: images and questions from V*Bench and HR-Bench; their terms apply.
• video-quality: videos from YouTube-UGC (CC BY 4.0) and Rapidata's Veo 3.1 human-preference data (Apache-2.0).
• edit-image-unlimited: 186 CC0 or public-domain images and 14 restoration photos provided by the task authors.
• egoexomem: questions from the EgoExoMem repository (CC BY 4.0).