-
Notifications
You must be signed in to change notification settings - Fork 0
/
notes.txt
32 lines (26 loc) · 1.48 KB
/
notes.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Features:
Show current coversation
Show past conversations
Ability to edit conversation tree
Can save conversation to file
Can send conversation to be trained
Keeps track of which conversations were already trained and what date and time
Can take a statement and generate prompts and completions resulting in a fine tune file
TODO:
Make it pretty
Allow for filtering conversations with tags
// Save conversations
tunes -> conversations -> tags/messages
upload tune w conversations
model formatting can be selected for any tune
option to create completions skipping some of the beginning messages, so that it doesn't look like the conversation has to start at a particular point.
Add token counts for selected model to conversation page
loras that keep sync of which conversations have been synced to the lora, and create a queue of conversations to be synced
What to do about updated conversations? Start over?
Continous mode which uses PPO algo to assign reward to outputs, can be edited to go back to normal fine tuning
Data from continous mode is saved for both conversation and reward
RLHI mode which trains discriminators to critique output using the same methods as normal tunes
A descriminator can be used to reinforce facts, ideas, principals, or personality quirks
Add ability to have bot respond as user
Add ability to create discriminators by a 5 star system and automatically build dataset.
regenerate a single message based on the messages before it.