THULab/cosmogony_multimodal_v3_scaled

COSMOGONY Multimodal v3 Scaled (TsFile)

Apache TsFile version of cmgyai/cosmogony-multimodal-v3-scaled.

Overview

COSMOGONY Multimodal v3 Scaled is a public, mission-aligned COSMOGONY training corpus for multimodal space-physics representation learning. It is curated from public providers and published as a materialized dataset for direct training. Each row is one observation: a stable identifier, a modality, a named provider and provenance URL, a time or catalog field, a per-row named feature vector, and a long task_relevance sentence describing the intended downstream reasoning.

• Rows: 4,446,916 (split train), published as 6 parquet shards and converted here into 6 .tsfile files with one shared schema.
• Modalities: space_weather_timeseries, orbital_elements, light_curve, planetary_system_catalog.
• Source families: NASA SPDF OMNI hourly products; NASA SPDF OMNI high-resolution minute products (2020–2026); CelesTrak selected public orbital groups; NASA Exoplanet Archive catalog products; NASA DONKI solar event records; GCAT satellite catalog; CelesTrak Consolidated Space Weather; NOAA/CelesTrak space-weather indices; Asteroid Lightcurve Database (LCDB); and an exoplanet-transit detection catalog (Kepler/K2/TESS).
• Time span: observations and catalog epochs from 1957 onward, plus a small set of rows whose source field is not a wall-clock instant (see notes below).

Schema (TsFile structure)

• Time (INT64, milliseconds) — epoch ms derived from the source timestamp column. Rows whose source timestamp is an empty string (50,118) or a bare BJD-like float (15,721) have no true instant; they are kept and assigned a deterministic synthetic placeholder starting at 2100-01-01T00:00:00Z.
• modality (TAG) — the source one-element modality list joined into a string. Example query: WHERE modality = 'space_weather_timeseries'.
• source (TAG) — provider name, e.g. WHERE source = 'NASA OMNI2 hourly'.
• task_relevance (TAG) — the long downstream-task sentence for the row.
• seq (TAG) — disambiguation tag added during conversion. Several catalog providers publish date-level records, so many rows share the same (modality, source, task_relevance, Time) key; TsFile does not allow duplicate timestamps inside one device, and seq is the per-key ordinal that keeps every row addressable without changing any source value.
• record_id (FIELD, STRING) — stable source identifier (unique per row).
• source_url (FIELD, STRING) — provenance URL exactly as published.
• timestamp_raw (FIELD, STRING) — the original timestamp string, kept verbatim so no information is lost when Time cannot be parsed.
• feature_names (FIELD, STRING) — JSON array naming the entries of features for that row.
• features_0 … features_19 (FIELD, DOUBLE) — the source features vector exploded into fixed columns.

Heterogeneous feature vectors. features is not homogeneous across the corpus: its length and meaning depend on modality/source, and feature_names records that meaning per row (observed lengths 7, 8, 10, 11, 12, 14, 16 and 20). To produce one consistent TsFile schema, every row was padded to the global maximum length of 20 before exploding, so features_0 … features_19 exist in every file. Shorter rows are NaN-padded (written as null) in the extra columns; nothing is truncated and no feature entry is dropped. Always read feature_names together with the row's modality/source before interpreting a features_i column.

FAIR-oriented curation notes

The original dataset follows repository guidelines for findability, accessibility, interoperability and reusability:

• Findable: each row has a stable record_id, explicit source, and source_url.
• Accessible: source access methods remain transparent via provider URLs.
• Interoperable: normalized schema and modality vocabulary across source families.
• Reusable: fixed provenance fields and explicit task-intent metadata (task_relevance).

Quality controls applied before publication include non-empty modality, aligned feature_names/features lengths, required identity/time fields, deterministic ordering by record_id, and duplicate record_id removal.

Intended uses

• self-supervised representation learning for space-physics signals
• masked reconstruction and forecasting pretraining
• multimodal encoder warm-start for downstream COSMOGONY models

Limitations

• This release is not a complete mirror of all upstream archives.
• License posture is source-specific, so the repository-level license remains other.
• Modal coverage is weighted toward space-weather time series in this version.
• Raw catalog records (empty or BJD-like timestamps) are preserved but their TsFile Time is a synthetic placeholder rather than a true observation time.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("cosmogony_multimodal_v3_scaled_part00.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], batch_size=1024) as result:
            batch = result.read_arrow_batch()
            if batch is not None:
                print(batch.to_pandas().head())

Source & license

• Original dataset: https://huggingface.co/datasets/cmgyai/cosmogony-multimodal-v3-scaled
• Author / publisher: COSMOGONY (cmgyai); curated from public space-physics providers
• License: other — licensing is source-specific. The original card declares repository-level license: other and each row keeps its own source / source_url provenance; please defer to the original providers for the terms of individual records.