seanpohorence/physim-worlds

Physim worlds

This snapshot contains 24 world records, 19 distinct genomes, 21 preserved records, and 3 eval-ready laboratory preparations. Multiple records may share a genome. The viewer's train label is a catalog partition, not a scientific training/test split. These are disclosed development data.

Catalog viewer and Parquet conversion

The worlds subset contains one row per registry record; evaluations contains one row per prepared evaluation bundle. Hugging Face automatically converts these two JSONL catalogs to Parquet on refs/convert/parquet to power browsing, filtering, and queries. That is a derived catalog view. Scientific arrays remain in their checksummed NPZ files, and the preserved registry records remain JSON.

Use a pinned dataset commit for scientific reproduction. The conversion branch follows the current catalog and is not the identity of a world or an evaluation.

from datasets import load_dataset

worlds = load_dataset("seanpohorence/physim-worlds", "worlds",
                      revision="FULL_DATASET_COMMIT", split="train")
evaluations = load_dataset("seanpohorence/physim-worlds", "evaluations",
                           revision="FULL_DATASET_COMMIT", split="train")

In worlds, world_record and genome are immutable identifiers; record_path and genome_path locate their JSON records. recipe and run link available provenance. A null bundle_path means the record has no prepared evaluation bundle. Paths are relative to this dataset at the selected commit. The evaluation catalog adds preparation/suite references, port counts, grading-program counts, download-profile sizes in bytes, dependencies, and licenses. Neither catalog loads or executes archived recipe code.

Worlds, preparations, and suites

• A world defines physical dynamics: its genome, parameters, and interactions. worlds.jsonl lists all records and links to their definitions and provenance.
• A laboratory preparation binds an exact saved physical starting state to its measurement/control apparatus. The apparatus belongs to this setup, not the world's intrinsic physics.
• An evaluation suite contains action/query programs, retained independent reference trajectories, and scoring configuration for a preparation.

One world can support multiple preparations and suites. preserved means its definition and available history are archived. eval-ready means an exact preparation and suite have been validated and packaged for evaluation. The Physim taskset defaults to every evaluation preparation at its pinned dataset revision; explicit selection can restrict a run to one preparation or a subset.

Contributing worlds

See the contribution guide for packaging a world with Blobkit and submitting an HF pull request without main-branch write access. Preserved worlds need a definition and available provenance; eval-ready preparations need the additional suite and validation evidence. Maintainers review submissions and regenerate the catalogs as part of publication.

Evaluation preparations

Preparation	Public ports	Grading programs	Files
bf_trail_lab_centered_v2	4	15	bundle
p4g2_044_centered_v2	12	19	bundle
xv_rotor_lab_centered_v2	6	15	bundle

catalog.jsonl lists these bundles, identities, profile sizes, and runtime requirements. Each program has two independent retained truth realizations; model evaluation requests 64 forecast samples. Grading programs within a bundle are separate from rollout counts.

Download and use

Select the full 40-character commit of this dataset to reproduce a snapshot. Replace FULL_DATASET_COMMIT and BUNDLE_PATH_FROM_CATALOG below; a mutable branch such as main is not accepted by the Physim downloader.

physim catalog --repo seanpohorence/physim-worlds --revision FULL_DATASET_COMMIT
physim fetch --repo seanpohorence/physim-worlds --revision FULL_DATASET_COMMIT \
  --path BUNDLE_PATH_FROM_CATALOG --profile evaluation

The fetch command returns a verified local directory. Supply it as env.taskset.task.tools.bundle to the environment. The simulation profile downloads the prepared laboratory without grading truths; evaluation includes them. Repeat fetch with --offline to use the verified cache.

To inspect general worlds and provenance, download the registry at the same dataset commit and use blobkit:

from huggingface_hub import snapshot_download
from blobkit.registry import Registry

snapshot = snapshot_download(
    "seanpohorence/physim-worlds", repo_type="dataset",
    revision="FULL_DATASET_COMMIT", allow_patterns=["registry/*"],
)
registry = Registry(snapshot + "/registry")
registry.verify()
world = registry.load_genome("WORLD_RECORD_ID_FROM_WORLDS_CATALOG")

The current preparations use Physim 0.13.0.dev0 and Blobkit 0.3.5. They bind a source centered on each device's sensor array; an injection remains at its launch position if the device subsequently moves. Superseded preparations are removed from the current catalog and file tree.

Use Python 3.12 with NumPy 2.5.2 and SciPy 1.18.0 on the trusted simulation host. The updated environment, installation instructions, and runnable configs are in the PhySim environment PR. release.json records the exact code source used for this dataset export. Blobkit is available as a checksummed release. The PyPI project named physim is unrelated.

Provenance and reproducibility

registry/index.json links immutable genomes, recipes, generation runs, candidate lineage, checkpoints, and harvested records. registry/availability.json links validated preparations and supporting evidence. Exact recipe sources and inputs, simulation observations, controls, and available pilot artifacts are preserved as content-addressed artifacts. Original descriptive paths are recorded in the JSON records. Use Registry.export_recipe to materialize a recipe for inspection and explicit execution; loading or verifying data never executes recipe code.

BF and XV's preparation inputs and recipes have been replay-checked. Their preserved bundles are the byte-exact evaluation inputs. Some older genomes lack their complete original evolutionary search history; their records explicitly retain those gaps. The snapshot does not invent missing provenance.

Each bundle manifest binds world, preparation, and suite identities, numerical source hashes, dependencies, and every payload hash. release.json inventories the snapshot's files and records exporter source hashes. A dirty code checkout is identified as such; its Git baseline is not claimed as an exact code release.

Evaluation boundary and scope

The evaluated model receives its public interface, observations, and its own files. The trusted evaluator keeps the bundle and grading truths outside the agent runtime. Recipe sources and archived pilot files are for research and reproduction, not agent inputs. These public development suites do not establish performance on undisclosed worlds. The original reference's primary score groups also omit some fine switch/recovery timing. New preparations or suites receive new content identities.

Licensing

World data are CC-BY-4.0 (LICENSE); archived code is Apache-2.0 (LICENSE-CODE). Preserve source attribution and the content identities when reusing material. Archived predictor artifacts can include pickle files: treat those as code-bearing research artifacts. Evaluation bundle arrays are NPZ loaded without pickle.