| # Datasheet for SimVerse |
|
|
| > **⚠️ Anonymized for double-blind review.** Author, organization, contact, citation, and repository fields are placeholders until the review concludes. |
|
|
| This datasheet follows the structure proposed by Gebru et al. (2018), *Datasheets for Datasets* (arXiv:1803.09010). It documents the SimVerse multimodal-LLM benchmark dataset. |
|
|
| --- |
|
|
| ## 1. Motivation |
|
|
| ### For what purpose was the dataset created? |
|
|
| SimVerse was created to evaluate multimodal large language models (LLMs) on **interactive simulation puzzles** — tasks that combine visual perception with multi-step reasoning under physical or geometric constraints. Existing multimodal benchmarks emphasize either single-step visual question answering or long-form free-text generation; SimVerse fills a gap by requiring structured, programmatically-verifiable outputs (placements, action sequences, command scripts) over five disjoint puzzle types that share a uniform prompt skeleton and output contract for cross-task comparability. |
|
|
| ### Who created the dataset and on behalf of which entity? |
|
|
| Anonymized for double-blind review. To be filled in post-acceptance. |
|
|
| ### Who funded the creation of the dataset? |
|
|
| Anonymized for double-blind review. |
|
|
| ### Any other comments? |
|
|
| The dataset is paired with an open-source code release (linked from the dataset card after the review period) that includes the generators, prompt builders, parsers, validators, and runner scripts for each task. |
|
|
| --- |
|
|
| ## 2. Composition |
|
|
| ### What do the instances that comprise the dataset represent? |
|
|
| Each instance is one **puzzle level** drawn from one of five tasks. A level bundles: |
| - A natural-language prompt (system + user) the benchmark presents to a model |
| - One or more visual assets (images, or in cutrope a short video) |
| - Structured fields describing the level's parameters (board size, segment lengths, target coordinates, etc.) |
| - A reference solution under a task-specific schema |
|
|
| | Config | Task | Records | Modality | Output schema | |
| |---|---|---:|---|---| |
| | `voi` | Text-VOI shape placement (XOR rasterization) | 600 | target image + per-shape images | `{placements:[{shape, angle, vertex, grid}]}` | |
| | `cube1` | Cube reconstruction (six-face map) | 502 | blank cross-net + path-imprint | `{faces:{TOP,BOTTOM,FRONT,BACK,LEFT,RIGHT}}` | |
| | `cube2` | Cube goal-roll (top-face directions) | 502 | initial cross-net + target image | `{directions:["N","S","E","W"]}` | |
| | `lamp` | Mechanical-arm lamp targeting | 610 | rendered workspace image | `{actions:[{joint, angle}]}` | |
| | `cutrope` | Cut the Rope video → command script | 272 | short MP4 gameplay clip | `{commands, reason, confidence}` | |
|
|
| **Total: 2,486 records.** |
|
|
| ### How many instances are there in total? |
|
|
| 2,486 records, broken down above. |
|
|
| ### Does the dataset contain all possible instances or is it a sample? |
|
|
| Each task has a programmatic generator that can produce arbitrarily many puzzles. The shipped 2,486 records are a **curated subset** sampled to span the difficulty distribution each generator's design parameters allow. The generators are publicly released so users can produce additional instances for held-out evaluation if they wish. |
|
|
| ### What data does each instance consist of? |
|
|
| Per-record JSON with fields including: |
| - `__sample_id__` — stable level id |
| - `prompt.system` — verbatim system-prompt text shown to the model |
| - `prompt.user` — verbatim user-prompt text (filled-in 9-section skeleton) |
| - `answer` — reference solution under the task-specific schema (see table above) |
| - task-native structured fields (e.g. `arm.segments`, `inventory`, `roll_sequence`) |
| - `images_relative_to_config` / `video_relative_to_config` — paths to media assets |
|
|
| Media files (PNG images, MP4 videos) live in sibling directories referenced by the JSON. |
|
|
| ### Is there a label or target associated with each instance? |
|
|
| Yes. The `answer` field carries a reference solution. For closed-form tasks (VOI, cube1, lamp) the answer is uniquely determined; for open-ended tasks (cube2, cutrope) the answer is one of many valid solutions and validators score by simulation outcome rather than string equality. |
|
|
| ### Is any information missing from individual instances? |
|
|
| cube1 deliberately includes `?` sentinel patterns when puzzle constraints under-determine a face. This is documented behavior, not missing data. |
|
|
| ### Are relationships between individual instances made explicit? |
|
|
| Each task's records are independent (no cross-record relationships). Records within a task share visual style, output schema, and prompt skeleton, but are otherwise independent puzzles. |
|
|
| ### Are there recommended data splits? |
|
|
| The dataset is **held-out evaluation only — there is no train split**. The HuggingFace YAML frontmatter declares a single `test` split per config. |
|
|
| ### Are there any errors, sources of noise, or redundancies in the dataset? |
|
|
| Records are programmatically validated by their respective engines before shipping, so every shipped record is provably solvable. Two known minor caveats: |
|
|
| 1. cube1's `?` sentinel pattern (intentional, documented). |
| 2. For open-ended tasks, the embedded reference solution is one of many valid solutions; downstream tooling that scores via reference equality rather than simulation will systematically miss credit for divergent strategies. |
|
|
| ### Is the dataset self-contained, or does it link to or otherwise rely on external resources? |
|
|
| The dataset is **self-contained**: every record carries the literal prompt, the reference answer, and either embedded paths to bundled media (images / videos) or all the structured fields needed to reproduce the task. No external API access, no external dataset dependencies, no licenced external content. |
|
|
| cutrope videos are derived from the open-source [`yell0wsuit/cuttherope-h5dx`](https://github.com/yell0wsuit/cuttherope-h5dx) port (MIT-licensed); the generated MP4 files are shipped within this dataset. |
|
|
| ### Does the dataset contain data that might be considered confidential? |
|
|
| No. Fully synthetic; no proprietary or confidential content. |
|
|
| ### Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety? |
|
|
| No. All visual content is abstract puzzle imagery (geometric shapes, cube nets, mechanical-arm renderings, gameplay clips of a children's puzzle game). |
|
|
| ### Does the dataset relate to people? |
|
|
| No. |
|
|
| ### Does the dataset identify any subpopulations? |
|
|
| No (the dataset does not contain people). |
|
|
| ### Is it possible to identify individuals, either directly or indirectly? |
|
|
| No. |
|
|
| ### Does the dataset contain data that might be considered sensitive? |
|
|
| No. |
|
|
| ### Any other comments? |
|
|
| None. |
|
|
| --- |
|
|
| ## 3. Collection Process |
|
|
| ### How was the data associated with each instance acquired? |
|
|
| All five tasks use **deterministic programmatic generators** (Python and JavaScript). Generators produce a level definition + a reference solution simultaneously, so the answer is not annotated post-hoc but co-produced with the puzzle. |
|
|
| - **VOI**: a polygon-rasterization XOR generator chooses base shapes, picks angles and vertices, and verifies the XOR-union matches a randomly-seeded target. |
| - **cube1 / cube2**: a Python cube-state simulator records roll sequences and bottom-face imprints. |
| - **lamp**: a forward-kinematics generator samples joint angles, places obstacles to constrain feasible solutions, and checks the bulb-target distance against the light radius. |
| - **cutrope**: built on top of the open-source `yell0wsuit/cuttherope-h5dx` HTML5 port; the SimVerse code records deterministic gameplay clips and authors matching command scripts. |
|
|
| Visual rendering is done by each task's frontend (HTML/JS) or a headless renderer. |
|
|
| ### What mechanisms or procedures were used to collect the data? |
|
|
| Generators were run as one-off scripts on author-provisioned commodity hardware. No human annotators, no crowdsourcing, no scraping. Determinism was enforced by fixed random seeds. |
|
|
| ### If the dataset is a sample from a larger set, what was the sampling strategy? |
|
|
| For each task, generators sweep difficulty parameters (move count, segment count, distractor count) and the shipped records are filtered to span the parameter ranges roughly uniformly within hand-tuned tier definitions. For example, cube1 records are bucketed into 5 difficulty tiers of ~100 records each. |
|
|
| ### Who was involved in the data collection process and how were they compensated? |
|
|
| The dataset's authors. No external contributors, no compensated annotators (anonymized for review). |
|
|
| ### Over what timeframe was the data collected? |
|
|
| The shipped subset was generated and curated during 2025–2026. |
|
|
| ### Were any ethical review processes conducted? |
|
|
| Not applicable — the dataset is fully synthetic and contains no human-subject data. |
|
|
| --- |
|
|
| ## 4. Preprocessing / cleaning / labeling |
|
|
| ### Was any preprocessing/cleaning/labeling of the data done? |
|
|
| Yes: |
| 1. **Schema migration** (`migrate_dataset.py` per task): legacy task-native answer fields were lifted into the unified v1 `answer` envelope (e.g. lamp's flat `[-60, -135, ...]` array → `{actions:[{joint, angle}]}`). |
| 2. **Prompt embedding** (`populate_prompts.py` per task): the literal system + user prompt strings the benchmark presents are embedded into each record so HF downloads need no auxiliary code. |
| 3. For cutrope, an additional `build_data.py` step precomputes per-level `prompt_level` metadata (object counts, canvas dimensions) from the source level files. |
|
|
| Both transformations are idempotent and reproducible from the accompanying code. |
|
|
| ### Was the "raw" data saved in addition to the preprocessed/cleaned/labeled data? |
|
|
| Yes — pre-migration formats are preserved in each record under `legacy_answer` (and for cutrope the original `reference_solution` string is kept alongside the v1 `answer`). These are scheduled for removal one release cycle after v1.0. |
|
|
| ### Is the software used to preprocess/clean/label the instances available? |
|
|
| Yes. The full preprocessing pipeline is open source in the accompanying repository (linked from the dataset card, post-acceptance). |
|
|
| --- |
|
|
| ## 5. Uses |
|
|
| ### Has the dataset been used for any tasks already? |
|
|
| It is being used for benchmark experiments described in the accompanying paper (anonymized). |
|
|
| ### Is there a repository that links to any or all papers or systems that use the dataset? |
|
|
| To be filled in post-acceptance. |
|
|
| ### What (other) tasks could the dataset be used for? |
|
|
| - Fine-grained ablations on visual chain-of-thought prompting strategies |
| - Studying cross-task transfer in multimodal LLMs |
| - Diagnostic evaluation of spatial reasoning, multi-step planning, or video understanding in isolation |
| - Probing whether a model's prompt-following improves with the FINAL_JSON output contract |
| |
| ### Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses? |
| |
| - **Synthetic visual style** per task: models that overfit to specific render styles may show inflated SimVerse scores that do not transfer to real-world spatial-reasoning tasks. Users should not interpret SimVerse performance as a real-world capability predictor. |
| - **Open-ended scoring**: cube2 and cutrope require simulation-based scoring; downstream uses that apply naive string-equality scoring will systematically under-credit divergent valid strategies. |
| - **English-only prompts**: the dataset does not test multilingual reasoning. |
| - **No train split**: do not use SimVerse for training; it is intended as held-out evaluation. |
| |
| ### Are there tasks for which the dataset should not be used? |
| |
| - **Should NOT** be used as training data for production multimodal LLMs. |
| - **Should NOT** be used as a stand-in for real-world spatial reasoning capability assessment. |
| - **Should NOT** be used to evaluate models on languages other than English. |
| |
| --- |
| |
| ## 6. Distribution |
| |
| ### Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created? |
| |
| Yes. The dataset is publicly distributed. |
| |
| ### How will the dataset will be distributed? |
| |
| Via Hugging Face Hub as a Dataset repository under the [Datasets API](https://huggingface.co/docs/datasets/) (one HuggingFace dataset, five configs). The dataset can also be downloaded directly with `huggingface-cli download` or via direct URLs. |
| |
| ### When will the dataset be distributed? |
| |
| The dataset will be made publicly available before the camera-ready deadline of the corresponding NeurIPS 2026 submission, in accordance with NeurIPS Datasets & Benchmarks track requirements. |
| |
| ### Will the dataset be distributed under a copyright or other intellectual property (IP) license, and/or under applicable terms of use (ToU)? |
| |
| Yes — **MIT License**. Full license text in `LICENSE` at the dataset root. |
| |
| ### Have any third parties imposed IP-based or other restrictions on the data associated with the instances? |
| |
| The cutrope subset is built on top of the [`yell0wsuit/cuttherope-h5dx`](https://github.com/yell0wsuit/cuttherope-h5dx) HTML5 port, which is itself MIT-licensed. The original Cut the Rope game IP is owned by ZeptoLab; the upstream port and SimVerse's derived gameplay clips are non-commercial educational/research uses consistent with the upstream MIT license. |
| |
| ### Do any export controls or other regulatory restrictions apply to the dataset or to individual instances? |
| |
| None known. |
| |
| --- |
| |
| ## 7. Maintenance |
| |
| ### Who will be supporting/hosting/maintaining the dataset? |
| |
| To be filled in post-acceptance. During review, the dataset is hosted as an anonymous account on Hugging Face Hub. |
| |
| ### How can the owner/curator/manager of the dataset be contacted? |
| |
| Anonymized for review. To be filled in post-acceptance. |
| |
| ### Is there an erratum? |
| |
| Not at this version. Errata, if any, will be tracked via the dataset's commit history on the Hugging Face Hub. |
| |
| ### Will the dataset be updated? |
| |
| The dataset uses semantic versioning. Patch updates fix data errors; minor updates add levels or fields; major updates change schema (with one release cycle of `legacy_*` field overlap). Updates are pushed as new tagged commits on Hugging Face. |
| |
| ### If the dataset relates to people, are there applicable limits on the retention of the data associated with the instances? |
| |
| Not applicable — no person data. |
| |
| ### Will older versions of the dataset continue to be supported/hosted/maintained? |
| |
| Yes. Hugging Face's commit history preserves all historical versions; users can pin to a specific revision via the `revision=` argument of `datasets.load_dataset(...)`. |
| |
| ### If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so? |
| |
| Yes. The accompanying open-source repository (linked post-acceptance) includes the full generator pipeline; contributors can fork, generate additional puzzles, and submit pull requests or publish derivative datasets. |
| |
| --- |
| |
| ## License |
| |
| MIT — see [LICENSE](LICENSE) at the dataset root. |
| |
