Is your feature request related to a problem? Please describe.
Currently, our agent code uses high-dimensional vectors from the vocab table for calculating reward vectors. We need to update the code to use the reduced embeddings from the token table.
Describe the solution you'd like
Identify and modify the classes and methods that use high-dimensional vectors.
Update the data loading process to fetch embeddings from the token table.
Adjust the reward calculation logic to use the reduced embeddings.
Update configuration and initialization routines to reflect the changes.
Thoroughly test and validate the updated agent code.
Describe alternatives you've considered
Continuing to use high-dimensional vectors, but this is inefficient and not scalable.
Using a hybrid approach, but it adds unnecessary complexity.
Additional context
This change is necessary to leverage the reduced embeddings efficiently and improve the scalability of our agent code. The token table will contain the reduced embeddings generated by the PCA reduction script (reduce.py).
Is your feature request related to a problem? Please describe. Currently, our agent code uses high-dimensional vectors from the
vocab
table for calculating reward vectors. We need to update the code to use the reduced embeddings from thetoken
table.Describe the solution you'd like
token
table.Describe alternatives you've considered
Additional context This change is necessary to leverage the reduced embeddings efficiently and improve the scalability of our agent code. The
token
table will contain the reduced embeddings generated by the PCA reduction script (reduce.py
).