Query regarding Embedding Normalization and Similarity Score Range for PE-AV model

[Preet-Sojitra]

Hi there,

I am using facebook/pe-av-large following the example code provided in the model card (using the dot product: audio_embeds @ visual_embeds.T).

I noticed that the resulting similarity scores often exceed 1.0 (e.g., I am seeing scores around 1.1). This suggests the embeddings are not L2-normalized by default.

1. Are the embeddings intended to be used as unnormalized dot products?
2. Is there a known range for these scores? I am trying to set a threshold to filter "good" vs "bad" pairs. Should I manually L2-normalize the embeddings to interpret them as Cosine Similarity (-1 to 1)?

Thanks!