The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
rows: int64
width: int64
max_q: int64
max_p: int64
pad_id: int64
dtype: string
layout: string
sources: list<item: struct<id: string, source_id: int64, start: int64, end: int64>>
child 0, item: struct<id: string, source_id: int64, start: int64, end: int64>
child 0, id: string
child 1, source_id: int64
child 2, start: int64
child 3, end: int64
models: struct<mine-embed-v4: struct<params: double, per_task: struct<MTOPDomainClassification: double, XNLI (... 2106 chars omitted)
child 0, mine-embed-v4: struct<params: double, per_task: struct<MTOPDomainClassification: double, XNLI: double, STS22.v2: do (... 413 chars omitted)
child 0, params: double
child 1, per_task: struct<MTOPDomainClassification: double, XNLI: double, STS22.v2: double, STS17: double, SIB200Cluste (... 379 chars omitted)
child 0, MTOPDomainClassification: double
child 1, XNLI: double
child 2, STS22.v2: double
child 3, STS17: double
child 4, SIB200ClusteringS2S: double
child 5, STSBenchmarkMultilingualSTS: double
child 6, XPQARetrieval: double
child 7, MTOPIntentClassification: double
child 8, MIRACLRetrievalHardNegatives: double
child 9, OpusparcusPC: double
child 10, AmazonCounterfactualClassification: double
child 11, Tatoeba: double
child 12, NTREXBitextMining: double
child 13, MultilingualSentimentClassification: double
ch
...
double
child 15, MassiveIntentClassification: double
child 16, MassiveScenarioClassification: double
run: struct<date: timestamp[s], tasks: list<item: string>, langs: list<item: string>, max_len: int64, bat (... 61 chars omitted)
child 0, date: timestamp[s]
child 1, tasks: list<item: string>
child 0, item: string
child 2, langs: list<item: string>
child 0, item: string
child 3, max_len: int64
child 4, batch: int64
child 5, precision: string
child 6, hardware: string
child 7, mteb: string
model: string
leaderboard: struct<rows: list<item: struct<model: string, params: double, avg: double, new: bool, STS: double, C (... 179 chars omitted)
child 0, rows: list<item: struct<model: string, params: double, avg: double, new: bool, STS: double, Classification (... 112 chars omitted)
child 0, item: struct<model: string, params: double, avg: double, new: bool, STS: double, Classification: double, P (... 100 chars omitted)
child 0, model: string
child 1, params: double
child 2, avg: double
child 3, new: bool
child 4, STS: double
child 5, Classification: double
child 6, PairClassification: double
child 7, BitextMining: double
child 8, Clustering: double
child 9, Retrieval: double
child 10, rank: int64
child 1, common: list<item: string>
child 0, item: string
child 2, notes: list<item: null>
child 0, item: null
to
{'model': Value('string'), 'models': {'mine-embed-v4': {'params': Value('float64'), 'per_task': {'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}, 'multilingual-e5-small': {'params': Value('float64'), 'per_task': {'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}, 'p
...
, 'MassiveScenarioClassification': Value('float64')}}, 'granite-embedding-107m-multilingual': {'params': Value('float64'), 'per_task': {'MintakaRetrieval': Value('float64'), 'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}}, 'leaderboard': {'rows': List({'model': Value('string'), 'params': Value('float64'), 'avg': Value('float64'), 'new': Value('bool'), 'STS': Value('float64'), 'Classification': Value('float64'), 'PairClassification': Value('float64'), 'BitextMining': Value('float64'), 'Clustering': Value('float64'), 'Retrieval': Value('float64'), 'rank': Value('int64')}), 'common': List(Value('string')), 'notes': List(Value('null'))}, 'run': {'date': Value('timestamp[s]'), 'tasks': List(Value('string')), 'langs': List(Value('string')), 'max_len': Value('int64'), 'batch': Value('int64'), 'precision': Value('string'), 'hardware': Value('string'), 'mteb': Value('string')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 274, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
rows: int64
width: int64
max_q: int64
max_p: int64
pad_id: int64
dtype: string
layout: string
sources: list<item: struct<id: string, source_id: int64, start: int64, end: int64>>
child 0, item: struct<id: string, source_id: int64, start: int64, end: int64>
child 0, id: string
child 1, source_id: int64
child 2, start: int64
child 3, end: int64
models: struct<mine-embed-v4: struct<params: double, per_task: struct<MTOPDomainClassification: double, XNLI (... 2106 chars omitted)
child 0, mine-embed-v4: struct<params: double, per_task: struct<MTOPDomainClassification: double, XNLI: double, STS22.v2: do (... 413 chars omitted)
child 0, params: double
child 1, per_task: struct<MTOPDomainClassification: double, XNLI: double, STS22.v2: double, STS17: double, SIB200Cluste (... 379 chars omitted)
child 0, MTOPDomainClassification: double
child 1, XNLI: double
child 2, STS22.v2: double
child 3, STS17: double
child 4, SIB200ClusteringS2S: double
child 5, STSBenchmarkMultilingualSTS: double
child 6, XPQARetrieval: double
child 7, MTOPIntentClassification: double
child 8, MIRACLRetrievalHardNegatives: double
child 9, OpusparcusPC: double
child 10, AmazonCounterfactualClassification: double
child 11, Tatoeba: double
child 12, NTREXBitextMining: double
child 13, MultilingualSentimentClassification: double
ch
...
double
child 15, MassiveIntentClassification: double
child 16, MassiveScenarioClassification: double
run: struct<date: timestamp[s], tasks: list<item: string>, langs: list<item: string>, max_len: int64, bat (... 61 chars omitted)
child 0, date: timestamp[s]
child 1, tasks: list<item: string>
child 0, item: string
child 2, langs: list<item: string>
child 0, item: string
child 3, max_len: int64
child 4, batch: int64
child 5, precision: string
child 6, hardware: string
child 7, mteb: string
model: string
leaderboard: struct<rows: list<item: struct<model: string, params: double, avg: double, new: bool, STS: double, C (... 179 chars omitted)
child 0, rows: list<item: struct<model: string, params: double, avg: double, new: bool, STS: double, Classification (... 112 chars omitted)
child 0, item: struct<model: string, params: double, avg: double, new: bool, STS: double, Classification: double, P (... 100 chars omitted)
child 0, model: string
child 1, params: double
child 2, avg: double
child 3, new: bool
child 4, STS: double
child 5, Classification: double
child 6, PairClassification: double
child 7, BitextMining: double
child 8, Clustering: double
child 9, Retrieval: double
child 10, rank: int64
child 1, common: list<item: string>
child 0, item: string
child 2, notes: list<item: null>
child 0, item: null
to
{'model': Value('string'), 'models': {'mine-embed-v4': {'params': Value('float64'), 'per_task': {'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}, 'multilingual-e5-small': {'params': Value('float64'), 'per_task': {'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}, 'p
...
, 'MassiveScenarioClassification': Value('float64')}}, 'granite-embedding-107m-multilingual': {'params': Value('float64'), 'per_task': {'MintakaRetrieval': Value('float64'), 'MTOPDomainClassification': Value('float64'), 'XNLI': Value('float64'), 'STS22.v2': Value('float64'), 'STS17': Value('float64'), 'SIB200ClusteringS2S': Value('float64'), 'STSBenchmarkMultilingualSTS': Value('float64'), 'XPQARetrieval': Value('float64'), 'MTOPIntentClassification': Value('float64'), 'MIRACLRetrievalHardNegatives': Value('float64'), 'OpusparcusPC': Value('float64'), 'AmazonCounterfactualClassification': Value('float64'), 'Tatoeba': Value('float64'), 'NTREXBitextMining': Value('float64'), 'MultilingualSentimentClassification': Value('float64'), 'MassiveIntentClassification': Value('float64'), 'MassiveScenarioClassification': Value('float64')}}}, 'leaderboard': {'rows': List({'model': Value('string'), 'params': Value('float64'), 'avg': Value('float64'), 'new': Value('bool'), 'STS': Value('float64'), 'Classification': Value('float64'), 'PairClassification': Value('float64'), 'BitextMining': Value('float64'), 'Clustering': Value('float64'), 'Retrieval': Value('float64'), 'rank': Value('int64')}), 'common': List(Value('string')), 'notes': List(Value('null'))}, 'run': {'date': Value('timestamp[s]'), 'tasks': List(Value('string')), 'langs': List(Value('string')), 'max_len': Value('int64'), 'batch': Value('int64'), 'precision': Value('string'), 'hardware': Value('string'), 'mteb': Value('string')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
mine-embed multilingual pairs: one .npy file
4,647,029 cleaned, deduplicated text pairs in 71 public sources for training multilingual embedding models, already
tokenized and packed into one NumPy file: mine_embed_v4_pairs.npy.
| File | What |
|---|---|
mine_embed_v4_pairs.npy |
int32 array, shape (4,647,029, 323): one row per pair |
pack_meta.json |
row layout and the row range of every source |
tokenizer/ |
the tokenizer the ids belong to (intfloat/multilingual-e5-small) |
mine_embed_pack.py |
60-line loader (needs only numpy) |
sources.json |
sources, licenses and attribution |
Row layout: [source_id, len_q, len_p, query ids (padded to 64), passage ids (padded to 256)], padding id 1.
Rows are grouped by source (each source is one contiguous slice), so batches can be built from a single source and in-batch negatives never mix languages.
About 0.63B tokens in total. Queries are cut to 64 tokens and passages to 256.
Code example
import json, numpy as np
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
repo = "MineAITechnology/mine-embed-v3-dataset"
path = snapshot_download(repo, repo_type="dataset") # about 6.0 GB
import sys; sys.path.insert(0, path)
from mine_embed_pack import PackedPairs
data = PackedPairs(f"{path}/mine_embed_v4_pairs.npy", f"{path}/pack_meta.json") # memory-mapped, instant
tok = AutoTokenizer.from_pretrained(f"{path}/tokenizer")
src, q_ids, p_ids = data.row(1234)
print(data.meta["sources"][src]["id"], "|", tok.decode(q_ids), "->", tok.decode(p_ids)[:120])
for batch in data.batches(batch_size=256, seed=0): # source-homogeneous, shuffled, padded numpy arrays
q_ids, q_mask, p_ids, p_mask = batch["q_ids"], batch["q_mask"], batch["p_ids"], batch["p_mask"]
break # feed these to your encoder (torch.from_numpy(...))
Plain numpy works too: a = np.load("mine_embed_v4_pairs.npy", mmap_mode="r"); a[i, 0] is the source id, a[i, 3:3+a[i,1]] the query ids.
The ids are only valid with the tokenizer in tokenizer/. Text is recovered with tok.decode (not byte-exact: whitespace is normalized).
Sources
| Source id | Hub repo | Config | Language | Task | Rows | License | Bucket |
|---|---|---|---|---|---|---|---|
| aya_dataset | CohereLabs/aya_dataset | multi | retrieval | 170,176 | apache-2.0 | A | |
| gooaq | sentence-transformers/gooaq | pair | en | retrieval | 250,000 | apache-2.0 (upstream) | A |
| natural_questions | sentence-transformers/natural-questions | pair | en | retrieval | 100,000 | cc-by-sa-3.0 (upstream) | B |
| squad_v2 | rajpurkar/squad_v2 | squad_v2 | en | retrieval | 86,759 | cc-by-sa-4.0 | B |
| tydiqa | google-research-datasets/tydiqa | secondary_task | multi | retrieval | 49,755 | apache-2.0 | A |
| wiki_ar | wikimedia/wikipedia | 20231101.ar | ar | retrieval | 80,000 | cc-by-sa-3.0+gfdl | B |
| wiki_bn | wikimedia/wikipedia | 20231101.bn | bn | retrieval | 40,000 | cc-by-sa-3.0+gfdl | B |
| wiki_cs | wikimedia/wikipedia | 20231101.cs | cs | retrieval | 40,000 | cc-by-sa-3.0+gfdl | B |
| wiki_de | wikimedia/wikipedia | 20231101.de | de | retrieval | 100,000 | cc-by-sa-3.0+gfdl | B |
| wiki_el | wikimedia/wikipedia | 20231101.el | el | retrieval | 30,000 | cc-by-sa-3.0+gfdl | B |
| wiki_en | wikimedia/wikipedia | 20231101.en | en | retrieval | 120,000 | cc-by-sa-3.0+gfdl | B |
| wiki_es | wikimedia/wikipedia | 20231101.es | es | retrieval | 100,000 | cc-by-sa-3.0+gfdl | B |
| wiki_fa | wikimedia/wikipedia | 20231101.fa | fa | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_fr | wikimedia/wikipedia | 20231101.fr | fr | retrieval | 100,000 | cc-by-sa-3.0+gfdl | B |
| wiki_he | wikimedia/wikipedia | 20231101.he | he | retrieval | 40,000 | cc-by-sa-3.0+gfdl | B |
| wiki_hi | wikimedia/wikipedia | 20231101.hi | hi | retrieval | 50,000 | cc-by-sa-3.0+gfdl | B |
| wiki_hu | wikimedia/wikipedia | 20231101.hu | hu | retrieval | 30,000 | cc-by-sa-3.0+gfdl | B |
| wiki_id | wikimedia/wikipedia | 20231101.id | id | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_it | wikimedia/wikipedia | 20231101.it | it | retrieval | 80,000 | cc-by-sa-3.0+gfdl | B |
| wiki_ja | wikimedia/wikipedia | 20231101.ja | ja | retrieval | 90,000 | cc-by-sa-3.0+gfdl | B |
| wiki_ko | wikimedia/wikipedia | 20231101.ko | ko | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_nl | wikimedia/wikipedia | 20231101.nl | nl | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_pl | wikimedia/wikipedia | 20231101.pl | pl | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_pt | wikimedia/wikipedia | 20231101.pt | pt | retrieval | 80,000 | cc-by-sa-3.0+gfdl | B |
| wiki_ro | wikimedia/wikipedia | 20231101.ro | ro | retrieval | 30,000 | cc-by-sa-3.0+gfdl | B |
| wiki_ru | wikimedia/wikipedia | 20231101.ru | ru | retrieval | 100,000 | cc-by-sa-3.0+gfdl | B |
| wiki_sv | wikimedia/wikipedia | 20231101.sv | sv | retrieval | 40,000 | cc-by-sa-3.0+gfdl | B |
| wiki_sw | wikimedia/wikipedia | 20231101.sw | sw | retrieval | 15,000 | cc-by-sa-3.0+gfdl | B |
| wiki_th | wikimedia/wikipedia | 20231101.th | th | retrieval | 40,000 | cc-by-sa-3.0+gfdl | B |
| wiki_tr | wikimedia/wikipedia | 20231101.tr | tr | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_uk | wikimedia/wikipedia | 20231101.uk | uk | retrieval | 50,000 | cc-by-sa-3.0+gfdl | B |
| wiki_ur | wikimedia/wikipedia | 20231101.ur | ur | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_vi | wikimedia/wikipedia | 20231101.vi | vi | retrieval | 60,000 | cc-by-sa-3.0+gfdl | B |
| wiki_zh | wikimedia/wikipedia | 20231101.zh | zh | retrieval | 100,000 | cc-by-sa-3.0+gfdl | B |
| wikimatrix_ar | sentence-transformers/parallel-sentences-wikimatrix | en-ar | ar | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_bg | sentence-transformers/parallel-sentences-wikimatrix | en-bg | bg | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ca | sentence-transformers/parallel-sentences-wikimatrix | en-ca | ca | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_cs | sentence-transformers/parallel-sentences-wikimatrix | en-cs | cs | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_da | sentence-transformers/parallel-sentences-wikimatrix | en-da | da | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_de | sentence-transformers/parallel-sentences-wikimatrix | en-de | de | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_el | sentence-transformers/parallel-sentences-wikimatrix | en-el | el | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_es | sentence-transformers/parallel-sentences-wikimatrix | en-es | es | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_et | sentence-transformers/parallel-sentences-wikimatrix | en-et | et | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_fa | sentence-transformers/parallel-sentences-wikimatrix | en-fa | fa | bitext | 39,524 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_fi | sentence-transformers/parallel-sentences-wikimatrix | en-fi | fi | bitext | 50,838 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_fr | sentence-transformers/parallel-sentences-wikimatrix | en-fr | fr | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_gl | sentence-transformers/parallel-sentences-wikimatrix | en-gl | gl | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_he | sentence-transformers/parallel-sentences-wikimatrix | en-he | he | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_hi | sentence-transformers/parallel-sentences-wikimatrix | en-hi | hi | bitext | 35,186 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_hr | sentence-transformers/parallel-sentences-wikimatrix | en-hr | hr | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_hu | sentence-transformers/parallel-sentences-wikimatrix | en-hu | hu | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_id | sentence-transformers/parallel-sentences-wikimatrix | en-id | id | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_it | sentence-transformers/parallel-sentences-wikimatrix | en-it | it | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ja | sentence-transformers/parallel-sentences-wikimatrix | en-ja | ja | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ka | sentence-transformers/parallel-sentences-wikimatrix | en-ka | ka | bitext | 999 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ko | sentence-transformers/parallel-sentences-wikimatrix | en-ko | ko | bitext | 34,321 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_lt | sentence-transformers/parallel-sentences-wikimatrix | en-lt | lt | bitext | 20,347 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_mk | sentence-transformers/parallel-sentences-wikimatrix | en-mk | mk | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_nl | sentence-transformers/parallel-sentences-wikimatrix | en-nl | nl | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_pl | sentence-transformers/parallel-sentences-wikimatrix | en-pl | pl | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_pt | sentence-transformers/parallel-sentences-wikimatrix | en-pt | pt | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ro | sentence-transformers/parallel-sentences-wikimatrix | en-ro | ro | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_ru | sentence-transformers/parallel-sentences-wikimatrix | en-ru | ru | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_sk | sentence-transformers/parallel-sentences-wikimatrix | en-sk | sk | bitext | 34,124 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_sl | sentence-transformers/parallel-sentences-wikimatrix | en-sl | sl | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_sr | sentence-transformers/parallel-sentences-wikimatrix | en-sr | sr | bitext | 40,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_sv | sentence-transformers/parallel-sentences-wikimatrix | en-sv | sv | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_tr | sentence-transformers/parallel-sentences-wikimatrix | en-tr | tr | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_uk | sentence-transformers/parallel-sentences-wikimatrix | en-uk | uk | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_vi | sentence-transformers/parallel-sentences-wikimatrix | en-vi | vi | bitext | 60,000 | cc-by-sa-4.0 (upstream) | B |
| wikimatrix_zh | sentence-transformers/parallel-sentences-wikimatrix | en-zh | zh | bitext | 80,000 | cc-by-sa-4.0 (upstream) | B |
Sources that could not be loaded in this build (skipped, not part of the data): wikimatrix_ms.
Licenses
Rows come from sources with permissive (bucket A) or share-alike (bucket B) terms. The pairs are derived from third-party text,
so each source's own license and attribution terms apply to its slice (sources.json). Share-alike sources (Wikipedia: CC BY-SA 3.0 / GFDL,
WikiMatrix and SQuAD: CC BY-SA, Natural Questions: CC BY-SA 3.0) mean that redistributing derived data usually needs the same share-alike terms.
Two license tags (WikiMatrix and Natural Questions) are taken from the upstream projects because the Hub cards carry no tag, so please
check them yourself. This card is not legal advice.
How it was built
Streamed from the Hub, rows cleaned (whitespace normalized, too-short rows and rows where query equals answer dropped, long passages cut to
2,000 characters, mostly non-English rows dropped for English-only sources), exact-deduplicated within each source on a hash of query + positive,
tokenized with the intfloat/multilingual-e5-small tokenizer and packed. No MTEB test set was used as a source and no task used for the benchmark was trained on.
Benchmark: a model trained on this data
A private model trained on it (6800 steps x batch 256 (about 1.7M pairs), lr 3e-05, symmetric InfoNCE (temperature 1/50), embedding distillation from BAAI/bge-m3 (weight 3.0), word embeddings frozen) is rank 2 of 4 among the models measured here (65.72 average) on 16 multilingual MTEB tasks x 21 languages
(details: MineAITechnology/mine-embed-v3-pilot). Our own runs, identical settings for every model.
| Rank | Model | Params (M) | Avg (16 tasks) | STS | Classification | PairClassification | BitextMining | Clustering | Retrieval |
|---|---|---|---|---|---|---|---|---|---|
| 1 | multilingual-e5-small | 117.7 | 66.05 | 71.75 | 66.72 | 80.85 | 74.59 | 34.07 | 48.16 |
| 2 | mine-embed-v4 | 117.7 | 65.72 | 70.35 | 65.73 | 82.77 | 78.62 | 40.40 | 41.45 |
| 3 | paraphrase-multilingual-MiniLM-L12-v2 | 117.7 | 62.42 | 72.65 | 64.61 | 87.29 | 60.26 | 41.26 | 28.34 |
| 4 | granite-embedding-107m-multilingual | 107.0 | 61.82 | 67.40 | 63.80 | 77.72 | 59.58 | 39.25 | 45.17 |
Held-out in-source retrieval (rank each query's answer among up to 2,000 answers of the same source, averaged over all sources): recall@1 0.965 -> 0.971, recall@10 0.991 -> 0.993, MRR 0.975 -> 0.980 (before -> after training). This is an easy in-distribution check, not comparable with MTEB scores.
Limitations
- No synthetic data, no mined hard negatives, no medical / legal / finance Q&A.
- Wikipedia title -> article pairs are weak query-passage pairs; quality varies by language.
- Pairs are cleaned automatically; some noise remains.
- Downloads last month
- 37

