=================================================================

XGLUE train/dev/test file format (* denotes a specific language):

=================================================================

QAM:

(1) xglue.qam.en.train: question \t answer passage \t answer label

(2) xglue.qam.*.dev: question \t answer passage \t answer label

(3) xglue.qam.*.test: question \t answer passage

QADSM:

(1) xglue.qadsm.en.train: query \t ad title \t ad description \t relevance label

(2) xglue.qadsm.*.dev: query \t ad title \t ad description \t relevance label

(3) xglue.qadsm.*.test: query \t ad title \t ad description

WPR:

(1) xglue.wpr.en.train: query \t web page title \t web page snippet \t relevance label

(2) xglue.wpr.*.dev: query \t web page title \t web page snippet \t relevance label

(3) xglue.wpr.*.test: query \t web page title \t web page snippet

NC:

(1) xglue.nc.en.train: query \t news title \t news body \t news category

(2) xglue.nc.*.dev: query \t news title \t news body \t news category

(3) xglue.nc.*.test: query \t news title \t news body

QG:

(1) xglue.qg.en.src.train: answer passage

(2) xglue.qg.en.tgt.train: labeled question

(3) xglue.qg.*.src.dev: answer passage

(4) xglue.qg.*.tgt.dev: labeled question

(5) xglue.qg.*.src.test: answer passage

NTG:

(1) xglue.ntg.en.src.train: news body

(2) xglue.ntg.en.tgt.train: labeled news title

(3) xglue.ntg.*.src.dev: news body

(4) xglue.ntg.*.tgt.dev: labeled news title

(5) xglue.ntg.*.src.test: news body

XNLI:

(1) multinli.train.en.tsv premise \t hypothesis \t label

(2) xnli.dev.tsv language \t gold_label \t sentence1_binary_parse \t sentence2_binary_parse \t sentence1_parse \t sentence2_parse \t sentence1 \t sentence2 \t promptID \t ...

(3) xnli.test.tsv language \t gold_label \t sentence1_binary_parse \t sentence2_binary_parse \t sentence1_parse \t sentence2_parse \t sentence1 \t sentence2 \t promptID \t ...

NER:

(sentences are separated by a blank line)

(1) en.train word \space label

(2) *.dev word \space label

(3) *.test word \space label

POS:

(sentences are separated by a blank line)

(1) en.train word \space label

(2) *.dev word \space label

(3) *.test word \space label

PAWSX:

(The first line is header: "id" \t "sentence1" \t "sentence2" \t "label")

(1) en/train.tsv id \t sentence1 \t sentence2 \t label

(2) */dev_2k.tsv id \t sentence1 \t sentence2 \t label

(3) */test_2k.tsv id \t sentence1 \t sentence2 \t label

MLQA:

(1) squad1.1/train-v1.1.json

{

"title": "...",

"paragraphs": [

{

"context": "...",

"qas": [

{

"answers": [

{

"answer_start": ...,

"text": "..."

}

],

"question": "...",

"id": "..."

}]

}

}

(2) MLQA/MLQA_v1/dev/dev-context-*-question-*.json same as train set

(3) MLQA/MLQA_v1/test/test-context-*-question-*.json same as train set

=================================================================

XGLUE prediction file format (* denotes a specific language):

=================================================================

(NOTES: All the tasks except MLQA and PAWS-X: The line number of each prediction file should be same as the input file. The i-th line of input corresponding to i-th line of output.

PAWS-X's input file has one additional header. MLQA uses JSON format.)

QAM:

QAM/*.prediction: predicted label

(Label should be in [0, 1])

QADSM:

QADSM/*.prediction: predicted label

(Label should be in ["Good", "Bad"])

WPR:

WPR/*.prediction: predicted label

(Label should be in [0, 1, 2, 3, 4], the meaning of every rating is presented by the dictionary: {"perfect":"4", "excellent":"3", "good":"2", "fair":"1", "bad":"0"})

NC:

NC/*.prediction: predicted label

(Label should be in ["foodanddrink", "sports", "news", "entertainment", "health", "video", "finance", "travel", "lifestyle", "autos"])

QG:

QG/*.prediction: predicted question

NTG:

NTG/*.prediction: predicted news title

XNLI:

XNLI/*.prediction: predicted label

(Label should be in ["neutral", "contradiction", "entailment"])

NER:

NER/*.prediction: predicted label of each word

(Label should be in ["B-LOC", "B-MISC", "B-ORG", "B-PER", "I-LOC", "I-MISC", "I-ORG", "I-PER", "O"])

POS:

POS/*.prediction: predicted label of each word

(Label should be in ["ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X"])

PAWS-X:

PAWSX/*.prediction: predicted label

(Label should be in [0, 1])

MLQA:

MLQA/*.prediction: {"qas_id1": "answer text 1", "qas_id2": "answer text 2"}