search

play4honor / discord_quote_bot

A bot that brings the "quote" feature to Discord.
2 stars 1 forks source link

Enable custom training for !misquote #94

Open synsypa opened 3 years ago

synsypa commented 3 years ago

We should produce an API to that allows users to:

  1. Scrape message history from the channel (maybe using fetchMessages() or another package that does channel exports)
  2. Format a training dataset for AuthorNet
    • This will likely require some usage of server.members to produce the correct mapping of Users to IDs
  3. Train a copy of AuthorNet
  4. Smooth integration with bot deployment

This would allow both for more automated retraining of the AuthorNet, and allow a theoretical other quotebot instance to use their own AuthorNet for !misquote

zhangchuck commented 3 years ago

Hm, initial reaction is that this is a leak of data from the server which is maybe not good for our customers.

zhangchuck commented 3 years ago

Second thought is that, once we have DynamoDB or some other DB setup, this may be easier to do.

synsypa commented 3 years ago

sorry, im abusing jargon. What i mean here by API is actually just like, some clean functions to do this, not actually a remote API. theoretical end user would still have to bring their own training rig, we'd just help them do the pre-processing and set up the training loop.

That is something like !get_train to build the training set and store it in s3 (probably would require DM confirmation, maybe require specific permissions) and then a set of functions outside the main bot for training an AuthorNet object. The initial motivator was to make updating our model more streamlined, but theoretically it could be packaged as a way for users to power their own !misquote