[BB2] FAQ #4172
[BB2] FAQ #4172
Conversation
| You'll need to do two things: | ||
|
|
||
| 1. Set `--knowledge-access-method search_only` | ||
| 2. Set `--query-generator-model-file zoo:sea/bart_sq_gen/model` |
There was a problem hiding this comment.
Is this one necessary? Isn't that the default atm?
There was a problem hiding this comment.
No, this is not the default. The default uses --knowledge-access-method classify, with the query generator from BB2 (trained to either search or retrieve from memory)
There was a problem hiding this comment.
Sorry, I was referring to --query-generator-model-file zoo:sea/bart_sq_gen/model. That can be default, right?
There was a problem hiding this comment.
At the moment it's not, the default is zoo:blenderbot2/query_generator/model, which determines whether to search or retrieve from memory
There was a problem hiding this comment.
Oh, yeah, I remember now.
| 1. Set `--knowledge-access-method search_only` | ||
| 2. Set `--query-generator-model-file zoo:sea/bart_sq_gen/model` | ||
|
|
||
| ### How can I train with gold documents provided? |
There was a problem hiding this comment.
Should we add a note here that what we mean by gold docs. Maybe to keep it general enough we could say something like this:
Gold documents are any set of documents that you need you retriever to surface. We add them to make sure the generator model sees (is conditioned on) a certain set of document. Example use cases are as follows:
1- Adding documents to the mix of retrieved that you are certain have useful knowledge that the model should use for generating response.
2- Ensure reproducibility between experiments for retrievers that have randomized responses. For example, if you wanted your model to see the same documents that was shown to crowdsourcing agent
(eg WizIntGoldDocRetrieverFiDAgent model with wizard of internet dataset).
There was a problem hiding this comment.
Great idea, i'll add that note
address mojtaba comments
Patch description
I've included a README for BB2 with FAQ that I've seen in our issues. Hopefully we can direct people towards this if it has the answer...