NatLee/sentiment-classification-dataset-bundle

Sentiment Classification Dataset Bundle

Six English sentiment and subjectivity datasets in one place, as plain Parquet: 135K MyAnimeList anime reviews (MALCoD) plus the classic IMDB, SST-2, MR, SUBJ and MPQA sets. Every config has a text column and an integer label with named classes.

For a benchmark of today's methods on all six (TF-IDF, sentence embeddings, fine-tuned ModernBERT and DeBERTa-v3, a local LLM), with fixed splits and seeds, see sentiment-bundle-benchmark.

Quick start

from datasets import load_dataset

imdb = load_dataset("NatLee/sentiment-classification-dataset-bundle", "aclImdb")
anime = load_dataset("NatLee/sentiment-classification-dataset-bundle", "myanimelist", split="train")

print(imdb["train"].features["label"].names)   # ['negative', 'positive']
print(anime[0]["rating"], anime[0]["label"])    # 10 1

Plain pandas works too: pd.read_parquet("hf://datasets/NatLee/sentiment-classification-dataset-bundle/MR/train.parquet").

Configs

Config	One example is	Labels	Rows
myanimelist	a full MyAnimeList review, 2006-2019	negative (rating 1-5) / positive (6-10)	train 135,203
aclImdb	an IMDB movie review (pre-processed, see below)	negative / positive	train 25,000, test 25,000
SST2	a movie-review sentence or phrase	negative / positive	train 76,961, validation 872, test 1,821
MR	a Rotten Tomatoes snippet	negative / positive	train 10,662
SUBJ	a movie sentence	subjective / objective	train 10,000
MPQA	a short opinion phrase from news text	negative / positive	train 10,606

myanimelist has more columns: rating (the reviewer's overall score, 1-10), review_id, work_id, work_name and post_time. label is rating >= 6; use rating if you want a different cut-off. Splitting by post_time (for example, train before 2018 and test on 2018-2019) gives a realistic "future reviews" test set.

What changed in this revision (November 2026)

The 2023 revision stored the data as Python pickles prepared for the GURA model. Re-reading them for the benchmark above turned up two problems with the data itself, so this revision rebuilds the bundle:

• Parquet instead of pickles. A pickle can run arbitrary code while it loads, and the Hub viewer cannot show it. Parquet has neither problem.
• IMDB: 25,000 generated reviews removed. The old aclImdb.pkl training set had 50,000 rows: for each class, the 12,500 real reviews followed by 12,500 reviews that GURA's data augmentation glued together from random word spans. The generated ones are all about the average length (a word-count spread of 4.8 words, against 85.8 for the real ones). The training split is now the official 25,000 reviews; the test split is unchanged.
• MyAnimeList: misaligned labels replaced. The old myanimelist-sts.pkl held 135,077 summarised reviews but 135,206 scores, and the scores drifted out of line with the text (a few rows at the start, about 119 rows by the end), so most texts carried another review's score. The new myanimelist config takes the full reviews from the MALCoD CSV the old file was built from, which has exactly 135,206 rows in the same order as the scores. Three rows with an impossible rating (0 or
  11. are dropped; runs of layout whitespace from the scrape are collapsed to single spaces.
• Same names, cleaner columns. Configs keep the old file names (aclImdb, SST2, MR, SUBJ, MPQA; myanimelist-sts became myanimelist because it no longer holds summaries). The text column is text, labels are integers with class names, and SST-2's dev split is now validation.

Nothing else changed: SST2, MR, SUBJ and MPQA keep the same rows in the same order, and the IMDB test split is identical.

The old files

The pickles are still in this repository's history at revision d1d7692228acbde386a054bcff9e3d0bda907fb3:

from huggingface_hub import hf_hub_download

path = hf_hub_download("NatLee/sentiment-classification-dataset-bundle", "MR.pkl", repo_type="dataset",
                       revision="d1d7692228acbde386a054bcff9e3d0bda907fb3")

Only unpickle them with a restricted loader (the benchmark ships one), and remember the two problems above.

Known quirks

• The aclImdb text was pre-processed in 2023: lower-cased, with punctuation and stop words removed, including not, no and never. Scores are a few points below the usual IMDB numbers because of it. For the raw reviews, use the original release.
• MR, SUBJ and MPQA have no official test split; everything is in train. Use cross-validation or a fixed seeded split.
• Duplicates are kept as published: MPQA repeats 2,365 phrases (for example "complaining") and SST2 398 phrases. De-duplicate before splitting if that matters to you.
• Two myanimelist reviews have empty text in the source. They are kept so row positions match the Kaggle CSV. About 81% of the reviews are positive, so report macro-F1 next to accuracy.

Sources and licences

The bundle collects datasets from their original publishers; each config keeps its own terms.

Config	Source	Licence
myanimelist	MALCoD on Kaggle (version 2), crawled from MyAnimeList.net	ODbL 1.0 (database), DbCL 1.0 (contents)
aclImdb	Large Movie Review Dataset	none stated; free for research, cite the paper
SST2	Stanford Sentiment Treebank	none stated; cite the paper
MR	sentence polarity dataset v1.0	none stated; cite the paper
SUBJ	subjectivity dataset v1.0	none stated; cite the paper
MPQA	MPQA opinion corpus	research and academic use only (the news texts)

Citation

N. Lee, "MyAnimeList Comment Dataset (MALCoD)," Kaggle, 2019.
A. L. Maas et al., "Learning Word Vectors for Sentiment Analysis," ACL 2011.
R. Socher et al., "Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank," EMNLP 2013.
B. Pang and L. Lee, "Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales," ACL 2005.
B. Pang and L. Lee, "A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts," ACL 2004.
J. Wiebe, T. Wilson and C. Cardie, "Annotating Expressions of Opinions and Emotions in Language," Language Resources and Evaluation, 2005.