Skip to contents

The OlmoEarth v1.3 embeddings (Ai2, published on Source Cooperative as a Zarr v3 store) quantize each 128-dimension annual vector to Int8 with the signed_power scheme, power 2: sign(q) * (|q| / 127.5)^2, with -128 as nodata. That is the decode dequantize_aef() applies too; this name records which product a pipeline reads. lazy_dataset() turns the -128 fill value into NaN at the read, so nodata stays NaN through the decode.

Usage

dequantize_olmoearth(x)

Arguments

x

Int8 codes (traced array or plain numeric).

Value

The dequantized values, same shape as x.

Details

Decode before any arithmetic: the quantization is nonlinear, so distances between raw codes are not proportional to distances between embeddings. The vectors are not unit length (the per-pixel norm over 128 dimensions is about 3.0 to 3.8); normalise them before taking a dot product as cosine similarity.

References

Team OlmoEarth et al. (2026). OlmoEarth v1.3 Embeddings. Ai2. https://allenai.org/papers/olmoearth-v1-3

https://source.coop/olmoearth/olmoearth-v1-3-embeddings