How chat models understand context from previous chat

[hushed yew]

hello, I have a seq2seq model which is trained on conversations to generate chat responses, but i want to make a model which can understand the context and can remember things.
for example

1. if I tell model what my name is, it must be able to recall my name after some messages
2. if I am talking about X and I asked some questions, then model must understand that my question is related to X and nothing else

[red shuttle]

Actually this is entity extraction and then saving entity for future use. You can create another util function to do the same and integrate with you seq2seq model.
As you are giving example name here, these kind of common entities can be detected by spacy ner so you can use that and can save the name into sqllite database or simply in the running memory if its just a demo project

But if you want to make this project scalable then I would suggest you to use RASA you can detect name as entity extraction and save it into memory using slots in RASA.

[hushed yew]

thanks mate

you know how can I implement RLHF on my gpt model?

for example

User: what is value of 2 + 2?
Bot: i don't know | Bot: 2 + 2 = 3 | Bot: 2 + 2 = 4
Rate: 0             Rate: 1          Rate: 2     

in which bot will generate some result

and user can rate it from 0 to 2