Is this model multi-class or multi-label?
The OpenAI moderation dataset is multi-label (many labels present in same promt)
But multi-class is mentioned here https://huggingface.co/KoalaAI/Text-Moderation#training-details
@KoalaAI could you please correct it or clarify?
Thanks
The dataset was modified to suit multi-class instead of multi-label. I simply duped entries in the dataset for each label they had, since that's what AutoTrain wanted & IIRC there was no option for multi-label.
The intended use for this classifier was for a in-game chat moderation, where it'd look at the top 3 labels and their scores and then decide based on that whether or not to block it; in reality I could have just made it "OK" and "NOT OK" for this purpose, but decided to keep the labels anyway in case someone might find it useful.
By duplicating entries and assigning only one label from the original multi-label set to each duplicate, it effectively transformed the problem. Each duplicate now represents a possible outcome in a multi-class setting. The strategy of selecting the top 3 labels from the multi-class model implicitly considers the original multi-label nature. Since an entry could appear with multiple labels in the transformed dataset, the labels with the highest probabilities are likely the most relevant ones.
I would like to eventually re-train this model (or a more advanced variant/successor) but so far I have't had the time as I am researching other subjects.
Let me know if you have any other questions or comments!