Hi @Amy_Bray :)

It seems there are some discrepancies in the test data (csv). A significant amount of the tweets are not in gold-random-json , but rather they are training examples in different versions of the dataset. For example, the tweet with id: ID_1001154804658286592 (What is happening to the infrastructure in New England...)is actually a training example with labels as shown here

I haven't looked at the training data but judging from the number of tweets (~76k) it does look like it stretches all the different versions of IDRIS (total is approx. 77.5k) and not just the gold-random-json (20k). May you confirm which dataset we're required to use. I hope you'll also look into the test set as well. I suggest we use the test_unlabeled.jsonl for which we don't have access to the true labels.