Recommend way to aggregate semantic code embeddings

[lazyhope]

What would be the recommend way to aggregate different semantic code embeddings from the same repository to represent the overall semantic of the repository? Currently I am averaging the UniXCoder embeddings and using cosine similarity score to evaluate the result, but I am not sure if it is the right way.

Thanks in advance!

[guoday]

If you lack supervised data to fine-tune UniXcoder, the averaging approach is appropriate.

[lazyhope]

If you lack supervised data to fine-tune UniXcoder, the averaging approach is appropriate.

Thank you!