Normalization of outputs
#2
by
skirres
- opened
First of all, thanks for the nice work and sharing your model and results 🤗
In the model card you mention the normalization of outputs, but in your repository I stumbled upon this line. I got two questions:
- Did you normalize the vectors for the retrieval tasks?
- And (also for the retrieval tasks) did you report the results of the
dotor thel2metric in the model card?
And just a tiny remark, if you always want to use normalization, you could consider specifying it in the configuration like for this model.
Hi,
- The default similarity for the retrieval task is cosine. So whether normalizing the embedding doesn't influence the results. This command only impacts the clustering and classification task.
- No. We compute the cosine similarity to retrieve relevant passages.
Thanks for your advice!🤗 We will consider to update the configuration.
We have updated the configuration. Thanks for your solution again!