Have you tried different prompts in bi-encode model

[ttjjlw]

hi Have you tried different prompts in bi-encode model. in your paper? i only find you tried different prompts in cross-encode model thanks

[Muennighoff]

for the bi-encoder i only tried the actual text itself surrounded by special tokens. I did not try adding additional text - I imagine that task-specific text like done in recent work could help. I tried a few different special tokens results of which are in the appendix I think & the brackets worked best