allenai/olmoearth_pretrain

What is OlmoEarth?

OlmoEarth is a family of multimodal, spatio‑temporal foundation models designed for Earth‑observation data. The models ingest satellite imagery (Sentinel‑2 optical, Sentinel‑1 radar, Landsat) together with derived geographic maps (e.g., OpenStreetMap, WorldCover, USDA Cropland, DEM, canopy height, WorldCereal) and learn a joint representation that can be used for downstream tasks such as segmentation, classification, or embedding generation.

The repository contains the pre‑training code and utilities for building, evaluating, and fine‑tuning these models. It is part of the broader OlmoEarth platform, which provides end‑to‑end tooling for planetary‑scale intelligence.

Key Features

Feature	Details
Model family	Nano, Tiny, Small, Base, Large – ranging from ~1 M to >300 M parameters (encoder + decoder).
Modalities	Sentinel‑2 (optical), Sentinel‑1 (SAR), Landsat, plus six derived map layers.
Training data	285 k samples (~2.56 km² each) drawn globally; dataset available on Hugging Face.
Versioning	v1, v1.1, v1.2 releases with incremental improvements; changelog linked.
Open‑source stack	Built on AllenAI’s OLMo‑core library; Docker images provided; optional flash‑attention support.
Licensing	OlmoEarth Artifact License (see `LICENSE`).
Documentation	Quick‑start for inference, fine‑tuning tutorials, dataset creation guide, evaluation scripts.

Getting Started

Installation (full stack)

# Clone the repo
git clone git@github.com:allenai/olmoearth_pretrain.git
cd olmoearth_pretrain

# Recommended: Python 3.12 + uv package manager
uv sync --locked --all-extras --python 3.12

The command creates an isolated virtual environment (.venv). Activate it with source .venv/bin/activate or run commands via uv run.

Inference‑only installation

If you only need to load a pretrained model (no training dependencies):

uv sync --locked

Using the Models

Load a checkpoint – model weights are hosted on Hugging Face (e.g., allenai/OlmoEarth-v1_2-Base).
Run inference – follow docs/Inference-Quickstart.md to instantiate the model and pass a satellite tile.
Fine‑tune – tutorials are provided for segmentation and embedding extraction:
- FinetuneOlmoEarthSegmentation.md
- OlmoEarthEmbeddings.md
- FinetuneOlmoEarth.md
Deploy – example deployments live in the companion olmoearth_projects repo.

Training Your Own Model

The repo includes ready‑made scripts under scripts/official/vX_X/ for each released version. Detailed steps are in docs/Pretraining.md, covering data preparation (docs/Dataset-Creation.md), environment setup, and launch commands.

Evaluation

Reproduce the authors’ benchmark results with the instructions in:

docs/Evaluation.md – research‑benchmark tests.
Partner‑task evaluations are documented in the rslearn_projects repository.

Testing

Run the test suites with uv run:

# Full suite (all optional deps except flash‑attn)
uv run --extra all-no-flash pytest tests/

# Model‑loading tests with full deps (includes olmo‑core)
uv run --extra all-no-flash pytest tests_minimal_deps/

# Minimal‑dependency loading tests (no olmo‑core)
uv run --extra dev pytest tests_minimal_deps/

The CI runs these twice to ensure the package works both with and without the heavy olmo-core dependency.

Citation

If you use the code (not just the model weights), cite the software:

@software{olmoearth_pretrain,
  author    = {{OlmoEarth Team}},
  title     = {{OlmoEarth}: A library for developing earth system foundation models},
  publisher = {Allen Institute for AI},
  url       = {https://github.com/allenai/olmoearth_pretrain},
  version   = {0.1.2},
  year      = {2026}
}

For a specific model, cite the corresponding technical report (links provided in the model table).

Who Might Use This?

Remote‑sensing researchers who need a pretrained backbone for satellite‑image tasks.
Geospatial data engineers building pipelines that require embeddings or segmentation masks.
AI labs interested in multimodal, spatio‑temporal foundation models and looking for a reference implementation.

Bottom Line

OlmoEarth‑pretrain is a fully‑featured, open‑source toolkit for training and deploying large‑scale Earth‑observation foundation models. It provides ready‑to‑use pretrained checkpoints, clear training scripts, and extensive documentation, making it a practical entry point for anyone working at the intersection of AI and planetary data.