Contents:
5 types of data will be used to train future iterations of finetuned models for data extraction:
- chat dataset with "na" output (chat input type)
- buy/sell dataset with "na" output (i.e. dataset not relevant for Rayman fist extraction) (advertisement input type)
- buy/sell dataset for Rayman fist with "DL" currency (advertisement input type)
- buy/sell dataset for Rayman fist with "BGL" currency (advertisement input type)
- buy/sell dataset for Rayman fist with "na" currency (advertisement input type)
This dataset was made after reviewing multiple finetuned models ranging from 135M to 1.7B params output the same wrong currency for user input that is not trained from the dataset i.e. an item with a price not mentioned in dataset, resulting in most models (except 3B v1.2 with thinking mode) to always output currency as "DL" instead of "BGL", and after reviewing the dataset, it turns out the ratio of DL and BGL mentioned in the dataset is 3:1, so the model is biased towards using DL as the currency most of the time regardless of the format logic provided.
Now, I have standardised every input type to have the same number of rows for (hopefully) less biased training and outputting values more accurately. Since the BGL dataset only have 150 rows, its the main limitting factor, so I had to set the dataset limit to 150 rows for the rest, which resulted in 750 rows in total.
One thing I have to mention, 4/5 of the dataset is classified under advertisement tag, so I'm going to expect the model to be more biased to responding the input_type as advertisement more often. If this hypothesis is true, then I will have to increase the chat data by 300 to balance out the classification data ratio (v0.1.1)
Also, I noticed that WL currency is denoted as a discord emoji with custom ID, which can affect model output quality post-training, so for next dataset iteration (v0.1.1) I have to replace the discord emoji for WL e.g. <:WL:1234> to just WL