Pre-extracted embeddings from 26 Earth-observation foundation models evaluated on 24 downstream tasks. The dataset was created by AI2 in 2025 to support the findings in the OlmoEarth paper. It contains model outputs for train, validation, and test splits using paper-best hyperparameters.
Use Cases
- Benchmark foundation model performance based on the 24 downstream tasks described.
- Compare embeddings from different Earth-observation models mentioned in the paper.
- Analyze model generalization using the provided train/validation/test splits.
- Reproduce or extend the evaluation results from the OlmoEarth paper.
Strengths
- Contains embeddings from 26 distinct Earth-observation foundation models.
- Covers evaluations across 24 downstream tasks as defined in the referenced paper.
- Includes outputs for train, validation, and test splits with specified hyperparameters.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- Allen Institute for AI (AI2)
- Collection Method
- Pre-extracted embeddings from model encoders run on task splits.
- Freshness
- Last updated 2026-05-15 05:45:17; freshness should be verified.