Embedded representations of popular textual datasets (IMDb, rotten tomatoes, yelp, ag-news, sst5, 20newsgroups, dbpedia, yahoo-answers, trec) for text quantification

Vectorial representations for common text classification datasets: Dataset Hugging Face ID #Classes imdb stanfordnlp/imdb 2 rt cornell-movie-review-data/rotten_tomatoes 2 yelp Yelp/yelp_review_full 5 ag_news fancyzhx/ag_news 4 sst5 SetFit/sst5 5 20newsgroups SetFit/20_newsgroups 20 dbpedia fancyzhx/dbpedia_14 14 yahoo_answers community-datasets/yahoo_answers_topics 10 trec SetFit/TREC-QC (coarse labels) 6 Extracted using the scripts in this repo, where more details of the extraction can be consulted. The embeddings are provided in different modalities: features: the next-to-last representation of the network logits: the pre-softmax values predictions: the post-softmax values (posterior probabilities) targets: the true labels Each file is stored in format npz (dictionary-like) and contains the numpy representations for the keys "train", "val", and "test" partitions for each dataset. The "train" partition was used to train the corresponding neural model, "val" was used for training monitoring, and "test" was not used during training. The name of the file uses the format <dataset>_<neural-model>_<modality>.npz; for example: "yelp_roberta-base_logits.npz". Currently, only roberta-base has been employed, but the extraction script allows other backbones be employed easily.