Table 2.

Data sets used in multi-task learning.

Data set	Task	Domain	Size	Example
STS-Ba	Sentence pair similarity	General	8600	Sentence 1: “A young child is riding a horse”; Sentence 2: “A child is riding a horse”; Similarity: 4.75
RQEb	Sentence pair classification	Biomedical	8900	Sentence 1: “Doctor X thinks he is probably just a normal 18 month old but would like to know if there are a certain number of respiratory infections that are considered normal for that age”; Sentence 2: “Probably a normal 18 month old but how many respiratory infections are normal”; Ground truth: entailment
MedNLIc	Sentence pair classification	Clinical	14,000	Sentence 1: “Labs were notable for Cr 1.7 (baseline 0.5 per old records) and lactate 2.4”; Sentence 2: “Patient has normal Cr”; Ground truth: contradiction
QQPd	Sentence pair classification	General	400,000	Sentence 1: “Why do rockets look white?”; Sentence 2: “Why are rockets and boosters painted white?”; Ground truth: 1
Topic	Sentence classification	Clinical	1,300,000	Sentence: “Negative for difficulty urinating, pain with urination, and frequent urination”; Ground truth: SIGNORSYMPTOM
MedNERe	Token-wise classification	Clinical	15,000	Sentence: “he developed respiratory distress on the AMf of admission, cough day PTAg, CXRh with B/Li LLj PNAk, started ciprofloxacin and levofloxacin”; Ground truth: ciprofloxacin [DRUG] levofloxacin [DRUG]

^aSTS-B: semantic textual similarity benchmark.

^bRQE: Recognizing Question Entailment.

^cMedNLI: natural language inference data set for the clinical domain.

^dQQP: Quora Question Pairs.

^eMedNER: medication named entity recognition.

^fAM: morning.

^gPTA: prior to admission.

^hCXR: chest x-ray.

ⁱB/L: bilateral.

^jLL: left lower.

^kPNA: pneumonia.