def _init():
subsets = {}
base_path = ir_datasets.util.home_path()/NAME
dlc = DownloadConfig.context(NAME, base_path)
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
collection = NytDocs(dlc['source'])
base = Dataset(collection, documentation('_'))
all_queries = NytQueries(collection)
all_qrels = NytQrels(collection)
match_qids = Lazy(lambda: VALID_IDS)
subsets['train'] = Dataset(
FilteredQueries(all_queries, match_qids, mode='exclude'),
FilteredQrels(all_qrels, match_qids, mode='exclude'),
collection,
documentation('train'))
subsets['valid'] = Dataset(
FilteredQueries(all_queries, match_qids, mode='include'),
FilteredQrels(all_qrels, match_qids, mode='include'),
collection,
documentation('valid'))
ir_datasets.registry.register('nyt', base)
for s in sorted(subsets):
ir_datasets.registry.register(f'nyt/{s}', subsets[s])
return base, subsets
def _init():
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
base_path = ir_datasets.util.home_path() / NAME
dlc = DownloadConfig.context(NAME, base_path, dua=DUA)
subsets = {}
collection = TrecDocs(dlc['docs'],
path_globs=[
'**/FBIS/FB*', '**/FR94/??/FR*', '**/FT/*/FT*',
'**/LATIMES/LA*'
],
namespace=NAME,
lang='en',
expected_file_count=2295,
count_hint=ir_datasets.util.count_hint(NAME))
queries = TrecQueries(GzipExtract(dlc['queries']),
namespace=NAME,
lang='en')
qrels = TrecQrels(dlc['qrels'], QREL_DEFS)
base = Dataset(collection, queries, qrels, documentation('_'))
for fold in FOLDS:
qid_filter = make_filter(fold)
subsets[fold] = Dataset(FilteredQueries(queries, qid_filter),
FilteredQrels(qrels, qid_filter), collection,
documentation(fold))
ir_datasets.registry.register(NAME, base)
for s in sorted(subsets):
ir_datasets.registry.register(f'{NAME}/{s}', subsets[s])
return base, subsets
def _init():
subsets = {}
base_path = ir_datasets.util.home_path()/NAME
dlc = DownloadConfig.context(NAME, base_path)
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
migrator = Migrator(base_path/'irds_version.txt', 'v2',
affected_files=[base_path/'nyt.tgz.pklz4'],
message='Migrating nyt (extracting body text)')
collection = migrator(NytDocs(dlc['source']))
base = Dataset(collection, documentation('_'))
# core17
subsets['trec-core-2017'] = Dataset(
TrecQueries(dlc['trec-core-2017/queries'], namespace='trec-core-2017', lang='en'),
TrecQrels(dlc['trec-core-2017/qrels'], CORE_QREL_DEFS),
collection,
documentation('trec-core-2017'))
# wksup
all_queries = NytQueries(collection)
all_qrels = NytQrels(collection)
match_qids = Lazy(lambda: VALID_IDS)
subsets['wksup'] = Dataset(
all_queries,
all_qrels,
collection,
documentation('wksup/train'))
subsets['wksup/train'] = Dataset(
FilteredQueries(all_queries, match_qids, mode='exclude'),
FilteredQrels(all_qrels, match_qids, mode='exclude'),
collection,
documentation('wksup/train'))
subsets['wksup/valid'] = Dataset(
FilteredQueries(all_queries, match_qids, mode='include'),
FilteredQrels(all_qrels, match_qids, mode='include'),
collection,
documentation('wksup/valid'))
ir_datasets.registry.register('nyt', base)
for s in sorted(subsets):
ir_datasets.registry.register(f'nyt/{s}', subsets[s])
return base, subsets
def _init():
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
base_path = ir_datasets.util.home_path() / NAME
dlc = DownloadConfig.context(NAME, base_path, dua=DUA)
migrator = Migrator(base_path / 'irds_version.txt',
'v2',
affected_files=[
base_path / 'collection.tsv',
base_path / 'collection.tsv.pklz4'
],
message=f'Migrating {NAME} (fixing passage encoding)')
collection = TsvDocs(Cache(
FixEncoding(TarExtract(dlc['collectionandqueries'], 'collection.tsv')),
base_path / 'collection.tsv'),
namespace='msmarco',
lang='en',
docstore_size_hint=14373971970,
count_hint=ir_datasets.util.count_hint(NAME))
collection = migrator(collection)
subsets = {}
subsets['train'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.train.tsv'),
base_path / 'train/queries.tsv'),
namespace='msmarco',
lang='en'),
TrecQrels(dlc['train/qrels'], QRELS_DEFS),
TsvDocPairs(GzipExtract(dlc['train/docpairs'])),
TrecScoredDocs(
Cache(
ExtractQidPid(
TarExtract(dlc['train/scoreddocs'], 'top1000.train.txt')),
base_path / 'train/ms.run')),
)
subsets['train/triples-v2'] = Dataset(
collection,
subsets['train'].queries_handler(),
subsets['train'].qrels_handler(),
TsvDocPairs(GzipExtract(dlc['train/docpairs/v2'])),
subsets['train'].scoreddocs_handler(),
)
subsets['train/triples-small'] = Dataset(
collection,
subsets['train'].queries_handler(),
subsets['train'].qrels_handler(),
TsvDocPairs(
Cache(
MapSmallTriplesQidPid(
TarExtract(dlc['train/docpairs/small'],
'triples.train.small.tsv'),
TarExtract(dlc['collectionandqueries'], 'collection.tsv'),
subsets['train'].queries_handler()),
base_path / 'train/small.triples.qidpid.tsv')),
subsets['train'].scoreddocs_handler(),
)
subsets['dev'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.dev.tsv'),
base_path / 'dev/queries.tsv'),
namespace='msmarco',
lang='en'),
TrecQrels(dlc['dev/qrels'], QRELS_DEFS),
)
subsets['dev/small'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(dlc['collectionandqueries'], 'queries.dev.small.tsv'),
base_path / 'dev/small/queries.tsv'),
namespace='msmarco',
lang='en'),
TrecQrels(
Cache(
TarExtract(dlc['collectionandqueries'], 'qrels.dev.small.tsv'),
base_path / 'dev/small/qrels'), QRELS_DEFS),
TrecScoredDocs(
Cache(
ExtractQidPid(TarExtract(dlc['dev/scoreddocs'],
'top1000.dev')),
base_path / 'dev/ms.run')),
)
subsets['eval'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.eval.tsv'),
base_path / 'eval/queries.tsv'),
namespace='msmarco',
lang='en'),
)
subsets['eval/small'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(dlc['collectionandqueries'], 'queries.eval.small.tsv'),
base_path / 'eval/small/queries.tsv'),
namespace='msmarco',
lang='en'),
TrecScoredDocs(
Cache(
ExtractQidPid(
TarExtract(dlc['eval/scoreddocs'], 'top1000.eval')),
base_path / 'eval/ms.run')),
)
subsets['trec-dl-2019'] = Dataset(
collection,
TrecQrels(dlc['trec-dl-2019/qrels'], TREC_DL_QRELS_DEFS),
TsvQueries(Cache(GzipExtract(dlc['trec-dl-2019/queries']),
base_path / 'trec-dl-2019/queries.tsv'),
namespace='msmarco',
lang='en'),
TrecScoredDocs(
Cache(ExtractQidPid(GzipExtract(dlc['trec-dl-2019/scoreddocs'])),
base_path / 'trec-dl-2019/ms.run')),
)
subsets['trec-dl-2020'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['trec-dl-2020/queries']),
namespace='msmarco',
lang='en'),
TrecQrels(dlc['trec-dl-2020/qrels'], TREC_DL_QRELS_DEFS),
TrecScoredDocs(
Cache(ExtractQidPid(GzipExtract(dlc['trec-dl-2020/scoreddocs'])),
base_path / 'trec-dl-2020/ms.run')),
)
# A few subsets that are contrainted to just the queries/qrels/docpairs that have at least
# 1 relevance assessment
train_judged = Lazy(
lambda: {q.query_id
for q in subsets['train'].qrels_iter()})
subsets['train/judged'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(), train_judged),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
train_judged),
subsets['train'],
)
dev_judged = Lazy(
lambda: {q.query_id
for q in subsets['dev'].qrels_iter()})
subsets['dev/judged'] = Dataset(
FilteredQueries(subsets['dev'].queries_handler(), dev_judged),
subsets['dev'],
)
dl19_judged = Lazy(
lambda: {q.query_id
for q in subsets['trec-dl-2019'].qrels_iter()})
subsets['trec-dl-2019/judged'] = Dataset(
FilteredQueries(subsets['trec-dl-2019'].queries_handler(),
dl19_judged),
FilteredScoredDocs(subsets['trec-dl-2019'].scoreddocs_handler(),
dl19_judged),
subsets['trec-dl-2019'],
)
dl20_judged = Lazy(
lambda: {q.query_id
for q in subsets['trec-dl-2020'].qrels_iter()})
subsets['trec-dl-2020/judged'] = Dataset(
FilteredQueries(subsets['trec-dl-2020'].queries_handler(),
dl20_judged),
FilteredScoredDocs(subsets['trec-dl-2020'].scoreddocs_handler(),
dl20_judged),
subsets['trec-dl-2020'],
)
# split200 -- 200 queries held out from the training data for validation
split200 = Lazy(lambda: SPLIT200_QIDS)
subsets['train/split200-train'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(),
split200,
mode='exclude'),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
split200,
mode='exclude'),
FilteredQrels(subsets['train'].qrels_handler(),
split200,
mode='exclude'),
FilteredDocPairs(subsets['train'].docpairs_handler(),
split200,
mode='exclude'),
subsets['train'],
)
subsets['train/split200-valid'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(),
split200,
mode='include'),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
split200,
mode='include'),
FilteredQrels(subsets['train'].qrels_handler(),
split200,
mode='include'),
FilteredDocPairs(subsets['train'].docpairs_handler(),
split200,
mode='include'),
subsets['train'],
)
# Medical subset
def train_med():
with dlc['medmarco_ids'].stream() as stream:
stream = codecs.getreader('utf8')(stream)
return {l.rstrip() for l in stream}
train_med = Lazy(train_med)
subsets['train/medical'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(), train_med),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(), train_med),
FilteredDocPairs(subsets['train'].docpairs_handler(), train_med),
FilteredQrels(subsets['train'].qrels_handler(), train_med),
subsets['train'],
)
# DL-Hard
dl_hard_qrels_migrator = Migrator(
base_path / 'trec-dl-hard' / 'irds_version.txt',
'v3',
affected_files=[base_path / 'trec-dl-hard' / 'qrels'],
message='Updating trec-dl-hard qrels')
hard_qids = Lazy(lambda: DL_HARD_QIDS)
dl_hard_base_queries = TsvQueries([
Cache(GzipExtract(dlc['trec-dl-2019/queries']),
base_path / 'trec-dl-2019/queries.tsv'),
Cache(GzipExtract(dlc['trec-dl-2020/queries']),
base_path / 'trec-dl-2020/queries.tsv')
],
namespace='msmarco',
lang='en')
subsets['trec-dl-hard'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
dl_hard_qrels_migrator(
TrecQrels(dlc['trec-dl-hard/qrels'], TREC_DL_QRELS_DEFS)),
documentation('trec-dl-hard'))
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['1'])
subsets['trec-dl-hard/fold1'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold1'))
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['2'])
subsets['trec-dl-hard/fold2'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold2'))
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['3'])
subsets['trec-dl-hard/fold3'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold3'))
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['4'])
subsets['trec-dl-hard/fold4'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold4'))
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['5'])
subsets['trec-dl-hard/fold5'] = Dataset(
collection, FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold5'))
ir_datasets.registry.register(NAME, Dataset(collection,
documentation('_')))
for s in sorted(subsets):
ir_datasets.registry.register(f'{NAME}/{s}',
Dataset(subsets[s], documentation(s)))
return collection, subsets
def _init():
base_path = ir_datasets.util.home_path() / NAME
dlc = DownloadConfig.context(NAME, base_path)
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
main_dlc = dlc['main']
collection = TsvDocs(Cache(
TarExtract(main_dlc, 'nfcorpus/raw/doc_dump.txt'),
base_path / 'collection.tsv'),
doc_cls=NfCorpusDoc,
namespace=NAME)
subsets = {}
def read_lines(file):
file = Cache(TarExtract(main_dlc, f'nfcorpus/raw/{file}'),
base_path / file)
with file.stream() as stream:
stream = codecs.getreader('utf8')(stream)
return {l.rstrip() for l in stream}
nontopic_qid_filter = Lazy(lambda: read_lines('nontopics.ids'))
video_qid_filter = Lazy(lambda: read_lines('all_videos.ids'))
subsets['train'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.titles.queries'),
base_path / 'train/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.all.queries'),
base_path / 'train/queries.all.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusQuery),
TrecQrels(
Cache(TarExtract(main_dlc, 'nfcorpus/train.3-2-1.qrel'),
base_path / 'train/qrels'), QRELS_DEFS),
documentation('train'),
)
subsets['train/nontopic'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.nontopic-titles.queries'),
base_path / 'train/nontopic/queries.tsv'),
namespace=NAME),
FilteredQrels(subsets['train'].qrels_handler(),
nontopic_qid_filter,
mode='include'),
documentation('train/nontopic'),
)
subsets['train/video'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.vid-titles.queries'),
base_path / 'train/video/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.vid-desc.queries'),
base_path / 'train/video/queries.desc.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusVideoQuery),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/train.nontopic-titles.queries'),
base_path / 'train/video/queries.tsv'),
NfCorpusVideoQuery,
namespace=NAME),
FilteredQrels(subsets['train'].qrels_handler(),
video_qid_filter,
mode='include'),
documentation('train/video'),
)
subsets['dev'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/dev.titles.queries'),
base_path / 'dev/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(TarExtract(main_dlc, 'nfcorpus/dev.all.queries'),
base_path / 'dev/queries.all.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusQuery),
TrecQrels(
Cache(TarExtract(main_dlc, 'nfcorpus/dev.3-2-1.qrel'),
base_path / 'dev/qrels'), QRELS_DEFS),
documentation('dev'),
)
subsets['dev/nontopic'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/dev.nontopic-titles.queries'),
base_path / 'dev/nontopic/queries.tsv'),
namespace=NAME),
FilteredQrels(subsets['dev'].qrels_handler(),
nontopic_qid_filter,
mode='include'),
documentation('dev/nontopic'),
)
subsets['dev/video'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/dev.vid-titles.queries'),
base_path / 'dev/video/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/dev.vid-desc.queries'),
base_path / 'dev/video/queries.desc.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusVideoQuery),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/dev.nontopic-titles.queries'),
base_path / 'dev/video/queries.tsv'),
NfCorpusVideoQuery,
namespace=NAME),
FilteredQrels(subsets['dev'].qrels_handler(),
video_qid_filter,
mode='include'),
documentation('dev/video'),
)
subsets['test'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/test.titles.queries'),
base_path / 'test/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(TarExtract(main_dlc, 'nfcorpus/test.all.queries'),
base_path / 'test/queries.all.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusQuery),
TrecQrels(
Cache(TarExtract(main_dlc, 'nfcorpus/test.3-2-1.qrel'),
base_path / 'test/qrels'), QRELS_DEFS),
documentation('test'),
)
subsets['test/nontopic'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/test.nontopic-titles.queries'),
base_path / 'test/nontopic/queries.tsv'),
namespace=NAME),
FilteredQrels(subsets['test'].qrels_handler(),
nontopic_qid_filter,
mode='include'),
documentation('test/nontopic'),
)
subsets['test/video'] = Dataset(
collection,
ZipQueries([
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/test.vid-titles.queries'),
base_path / 'test/video/queries.titles.tsv'),
namespace=NAME),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/test.vid-desc.queries'),
base_path / 'test/video/queries.desc.tsv'),
namespace=NAME),
], [(0, 0), (0, 1), (1, 1)], NfCorpusVideoQuery),
TsvQueries(Cache(
TarExtract(main_dlc, 'nfcorpus/test.nontopic-titles.queries'),
base_path / 'test/video/queries.tsv'),
NfCorpusVideoQuery,
namespace=NAME),
FilteredQrels(subsets['test'].qrels_handler(),
video_qid_filter,
mode='include'),
documentation('test/video'),
)
ir_datasets.registry.register(NAME, Dataset(collection,
documentation('_')))
for s in sorted(subsets):
ir_datasets.registry.register(f'{NAME}/{s}', subsets[s])
return collection, subsets
def _init():
base_path = ir_datasets.util.home_path()/NAME
documentation = YamlDocumentation(f'docs/{NAME}.yaml')
dlc = DownloadConfig.context(NAME, base_path, dua=DUA)
subsets = {}
collection = MsMarcoTrecDocs(GzipExtract(dlc['docs']))
subsets['train'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['train/queries']), namespace='msmarco', lang='en'),
TrecQrels(GzipExtract(dlc['train/qrels']), QRELS_DEFS),
TrecScoredDocs(GzipExtract(dlc['train/scoreddocs'])),
)
subsets['dev'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['dev/queries']), namespace='msmarco', lang='en'),
TrecQrels(GzipExtract(dlc['dev/qrels']), QRELS_DEFS),
TrecScoredDocs(GzipExtract(dlc['dev/scoreddocs'])),
)
subsets['eval'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['eval/queries']), namespace='msmarco', lang='en'),
TrecScoredDocs(GzipExtract(dlc['eval/scoreddocs'])),
)
subsets['trec-dl-2019'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['trec-dl-2019/queries']), namespace='msmarco', lang='en'),
TrecQrels(dlc['trec-dl-2019/qrels'], TREC_DL_QRELS_DEFS),
TrecScoredDocs(GzipExtract(dlc['trec-dl-2019/scoreddocs'])),
)
subsets['trec-dl-2020'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['trec-dl-2020/queries']), namespace='msmarco', lang='en'),
TrecQrels(dlc['trec-dl-2020/qrels'], TREC_DL_QRELS_DEFS),
TrecScoredDocs(GzipExtract(dlc['trec-dl-2020/scoreddocs'])),
)
subsets['orcas'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['orcas/queries']), namespace='orcas', lang='en'),
TrecQrels(GzipExtract(dlc['orcas/qrels']), ORCAS_QLRES_DEFS),
TrecScoredDocs(GzipExtract(dlc['orcas/scoreddocs'])),
)
dl19_judged = Lazy(lambda: {q.query_id for q in subsets['trec-dl-2019'].qrels_iter()})
subsets['trec-dl-2019/judged'] = Dataset(
FilteredQueries(subsets['trec-dl-2019'].queries_handler(), dl19_judged),
FilteredScoredDocs(subsets['trec-dl-2019'].scoreddocs_handler(), dl19_judged),
subsets['trec-dl-2019'],
)
dl20_judged = Lazy(lambda: {q.query_id for q in subsets['trec-dl-2020'].qrels_iter()})
subsets['trec-dl-2020/judged'] = Dataset(
FilteredQueries(subsets['trec-dl-2020'].queries_handler(), dl20_judged),
FilteredScoredDocs(subsets['trec-dl-2020'].scoreddocs_handler(), dl20_judged),
subsets['trec-dl-2020'],
)
# DL-Hard
dl_hard_qrels_migrator = Migrator(base_path/'trec-dl-hard'/'irds_version.txt', 'v2',
affected_files=[base_path/'trec-dl-hard'/'qrels'],
message='Updating trec-dl-hard qrels')
hard_qids = Lazy(lambda: DL_HARD_QIDS)
dl_hard_base_queries = TsvQueries([
Cache(GzipExtract(dlc['trec-dl-2019/queries']), base_path/'trec-dl-2019/queries.tsv'),
Cache(GzipExtract(dlc['trec-dl-2020/queries']), base_path/'trec-dl-2020/queries.tsv')], namespace='msmarco', lang='en')
subsets['trec-dl-hard'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
dl_hard_qrels_migrator(TrecQrels(dlc['trec-dl-hard/qrels'], TREC_DL_QRELS_DEFS)),
documentation('trec-dl-hard')
)
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['1'])
subsets['trec-dl-hard/fold1'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold1')
)
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['2'])
subsets['trec-dl-hard/fold2'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold2')
)
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['3'])
subsets['trec-dl-hard/fold3'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold3')
)
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['4'])
subsets['trec-dl-hard/fold4'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold4')
)
hard_qids = Lazy(lambda: DL_HARD_QIDS_BYFOLD['5'])
subsets['trec-dl-hard/fold5'] = Dataset(
collection,
FilteredQueries(dl_hard_base_queries, hard_qids),
FilteredQrels(subsets['trec-dl-hard'], hard_qids),
documentation('trec-dl-hard/fold5')
)
subsets['anchor-text'] = Dataset(
MsMarcoAnchorTextDocs(
Cache(GzipExtract(dlc['anchor-text']), base_path / "anchor-text.json"),
count_hint=1703834
),
documentation('anchor-text')
)
ir_datasets.registry.register(NAME, Dataset(collection, documentation("_")))
for s in sorted(subsets):
ir_datasets.registry.register(f'{NAME}/{s}', Dataset(subsets[s], documentation(s)))
return collection, subsets
def _init():
documentation = YamlDocumentation('docs/msmarco-passage.yaml')
base_path = ir_datasets.util.home_path() / 'msmarco-passage'
dlc = DownloadConfig.context('msmarco-passage', base_path, dua=DUA)
collection = TsvDocs(Cache(
FixEncoding(TarExtract(dlc['collectionandqueries'], 'collection.tsv')),
base_path / 'collection.tsv'),
namespace='msmarco')
subsets = {}
subsets['train'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.train.tsv'),
base_path / 'train/queries.tsv'),
namespace='msmarco'),
TrecQrels(dlc['train/qrels'], QRELS_DEFS),
TsvDocPairs(GzipExtract(dlc['train/docpairs'])),
TrecScoredDocs(
Cache(
ExtractQidPid(
TarExtract(dlc['train/scoreddocs'], 'top1000.train.txt')),
base_path / 'train/ms.run')),
)
subsets['dev'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.dev.tsv'),
base_path / 'dev/queries.tsv'),
namespace='msmarco'),
TrecQrels(dlc['dev/qrels'], QRELS_DEFS),
TrecScoredDocs(
Cache(
ExtractQidPid(TarExtract(dlc['dev/scoreddocs'],
'top1000.dev')),
base_path / 'dev/ms.run')),
)
subsets['dev/small'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(dlc['collectionandqueries'], 'queries.dev.small.tsv'),
base_path / 'dev/small/queries.tsv'),
namespace='msmarco'),
TrecQrels(
Cache(
TarExtract(dlc['collectionandqueries'], 'qrels.dev.small.tsv'),
base_path / 'dev/small/qrels'), QRELS_DEFS),
)
subsets['eval'] = Dataset(
collection,
TsvQueries(Cache(TarExtract(dlc['queries'], 'queries.eval.tsv'),
base_path / 'eval/queries.tsv'),
namespace='msmarco'),
TrecScoredDocs(
Cache(
ExtractQidPid(
TarExtract(dlc['eval/scoreddocs'], 'top1000.eval')),
base_path / 'eval/ms.run')),
)
subsets['eval/small'] = Dataset(
collection,
TsvQueries(Cache(
TarExtract(dlc['collectionandqueries'], 'queries.eval.small.tsv'),
base_path / 'eval/small/queries.tsv'),
namespace='msmarco'),
)
subsets['trec-dl-2019'] = Dataset(
collection,
TrecQrels(dlc['trec-dl-2019/qrels'], TREC_DL_QRELS_DEFS),
TsvQueries(Cache(GzipExtract(dlc['trec-dl-2019/queries']),
base_path / 'trec-dl-2019/queries.tsv'),
namespace='msmarco'),
TrecScoredDocs(
Cache(ExtractQidPid(GzipExtract(dlc['trec-dl-2019/scoreddocs'])),
base_path / 'trec-dl-2019/ms.run')),
)
subsets['trec-dl-2020'] = Dataset(
collection,
TsvQueries(GzipExtract(dlc['trec-dl-2020/queries']),
namespace='msmarco'),
TrecScoredDocs(
Cache(ExtractQidPid(GzipExtract(dlc['trec-dl-2020/scoreddocs'])),
base_path / 'trec-dl-2020/ms.run')),
)
# A few subsets that are contrainted to just the queries/qrels/docpairs that have at least
# 1 relevance assessment
train_judged = Lazy(
lambda: {q.query_id
for q in subsets['train'].qrels_iter()})
subsets['train/judged'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(), train_judged),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
train_judged),
subsets['train'],
)
dev_judged = Lazy(
lambda: {q.query_id
for q in subsets['dev'].qrels_iter()})
subsets['dev/judged'] = Dataset(
FilteredQueries(subsets['dev'].queries_handler(), dev_judged),
FilteredScoredDocs(subsets['dev'].scoreddocs_handler(), dev_judged),
subsets['dev'],
)
dl19_judged = Lazy(
lambda: {q.query_id
for q in subsets['trec-dl-2019'].qrels_iter()})
subsets['trec-dl-2019/judged'] = Dataset(
FilteredQueries(subsets['trec-dl-2019'].queries_handler(),
dl19_judged),
FilteredScoredDocs(subsets['trec-dl-2019'].scoreddocs_handler(),
dl19_judged),
subsets['trec-dl-2019'],
)
# split200 -- 200 queries held out from the training data for validation
split200 = Lazy(lambda: SPLIT200_QIDS)
subsets['train/split200-train'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(),
split200,
mode='exclude'),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
split200,
mode='exclude'),
FilteredQrels(subsets['train'].qrels_handler(),
split200,
mode='exclude'),
FilteredDocPairs(subsets['train'].docpairs_handler(),
split200,
mode='exclude'),
subsets['train'],
)
subsets['train/split200-valid'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(),
split200,
mode='include'),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(),
split200,
mode='include'),
FilteredQrels(subsets['train'].qrels_handler(),
split200,
mode='include'),
FilteredDocPairs(subsets['train'].docpairs_handler(),
split200,
mode='include'),
subsets['train'],
)
# Medical subset
def train_med():
with dlc['medmarco_ids'].stream() as stream:
stream = codecs.getreader('utf8')(stream)
return {l.rstrip() for l in stream}
train_med = Lazy(train_med)
subsets['train/medical'] = Dataset(
FilteredQueries(subsets['train'].queries_handler(), train_med),
FilteredScoredDocs(subsets['train'].scoreddocs_handler(), train_med),
FilteredDocPairs(subsets['train'].docpairs_handler(), train_med),
FilteredQrels(subsets['train'].qrels_handler(), train_med),
subsets['train'],
)
ir_datasets.registry.register('msmarco-passage',
Dataset(collection, documentation('_')))
for s in sorted(subsets):
ir_datasets.registry.register(f'msmarco-passage/{s}',
Dataset(subsets[s], documentation(s)))
return collection, subsets