Post
2793
It turned out that the following simple method seems to be actually effective when you want to increase the appearance probability of only one or a very limited number of tokens.
By training LoRA with unsloth based on the .txt file generated by the code above, you can increase the appearance probability of specific tokens while maintaining the model's performance to great extent. However, it's better to stop the training before train loss becomes 0.0, as it will start spamming the token once it appears even once. In general, you can stop training at a very early stage and it will still work.
It is also possible to reduce the appearance probability of specific tokens by creating an over-learned LoRA with the specific tokens you want to reduce, combining it with the model, and then creating a model that extracts only the difference using the chat vector method and subtracting it from an arbitrary model.
In this case, it is better to set the ratio of chat vector to about five times. It has very little effect on the overall performance, apart from the specific tokens.
import os
one_token = "♡" # Token to increase the appearance probability
value = 1000000
token = one_token * value
with open("one-token.txt", "w", encoding="utf-8") as f:
f.write(token)By training LoRA with unsloth based on the .txt file generated by the code above, you can increase the appearance probability of specific tokens while maintaining the model's performance to great extent. However, it's better to stop the training before train loss becomes 0.0, as it will start spamming the token once it appears even once. In general, you can stop training at a very early stage and it will still work.
It is also possible to reduce the appearance probability of specific tokens by creating an over-learned LoRA with the specific tokens you want to reduce, combining it with the model, and then creating a model that extracts only the difference using the chat vector method and subtracting it from an arbitrary model.
In this case, it is better to set the ratio of chat vector to about five times. It has very little effect on the overall performance, apart from the specific tokens.
new_v = v - (5.0 * chat_vector[i].to(v.device))