Differences

This shows you the differences between two versions of the page.

--- nlp:datasets [2023/05/19 09:44] – [NLP Datasets] jmflanig
+++ nlp:datasets [2023/11/29 21:14] (current) – jmflanig
@@ Line 8: / Line 8: @@
   * [[https://arxiv.org/pdf/1506.06724.pdf|Bookcorpus]] (Used in BERT)
-===== Multi-Task =====
+===== General Benchmarks or Multi-Task Benchmarks =====
   * [[https://gluebenchmark.com/|GLUE]]: [[https://arxiv.org/pdf/1804.07461.pdf|paper]] - Warning: Has an issue with QQP and WNLI due to dev and tests sets not coming from the same distribution. See FAQ [[https://gluebenchmark.com/faq/|here]].
   * [[https://super.gluebenchmark.com/|SuperGLUE]]: [[https://arxiv.org/pdf/1905.00537.pdf|paper]] - A more difficult version of GLUE.
   * [[https://www.cluebenchmarks.com/en/index.html|CLUE]]: [[https://arxiv.org/pdf/2004.05986.pdf|paper]] - Like GLUE, but for Chinese
+  * [[https://github.com/hendrycks/test|MMMLU]]: [[https://arxiv.org/pdf/2009.03300.pdf|Hendrycks et al 2020 - Measuring Massive Multitask Language Understanding]] This dataset is a popular dataset for LLMs to evaluate on (for example GPT-4, etc).  However, it has two serious issues. 1) the test set is available on the web, which means LLMs are likely contaminated, and 2) the datasets has no in-domain training data, and can only be evaluated in a few-shot manner. This make is impossible to properly compare to prior fine-tuned methods.
 ===== Multilingual =====