msmarco-scores-ms-marco-MiniLM-L6-v2

# MS MARCO query-passage scores using [cross-encoder/ms-marco-MiniLM-L6-v2](https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2)

[MS MARCO](https://microsoft.github.io/msmarco/) is a large scale information retrieval corpus that was created based on real user search queries using the Bing search engine.
This dataset contains 160 million CrossEncoder scores on the MS MARCO dataset, using the [cross-encoder/ms-marco-MiniLM-L6-v2](https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2) model.
The scores are unprocessed logits, i.e. they don't range between 0...1, and they can be used for finetuning search models using distillation.

See also the [MS MARCO Mined Triplets collection](https://huggingface.co/collections/sentence-transformers/ms-marco-mined-triplets-6644d6f1ff58c5103fe65f23) for triplets mined using 13 different embedding models, perhaps filtered using this dataset to avoid false negatives.

## Dataset Subsets

### `pair` subset

* Columns: "query_id", "corpus_id", "score"
* Column types: `int`, `int`, `float`
* Examples:
```python
{
"query_id": 571018,
"corpus_id": 7349777,
"score": 10.257502,
}
```
* Collection Strategy: Mining negatives using an embedding model, and rescoring the query/answer pairs. This subset contains all scored query-passage pairs.

### `triplet` subset

* Columns: "query_id", "positive_id", "negative_id", "score"
* Column types: `int`, `int`, `list[float]`
* Examples:
```python
{
"query_id": 571018,
"positive_id": 1283525,
"negative_id": 7349777,
"score": [3.421323299407959, 10.257501602172852],
}
```
* Collection Strategy: Mining negatives using an embedding model, and rescoring the query/answer pairs. Randomly subdivide all query/answer pairs for each query into two equally sized groups, the first group representing positives and the second group negatives. The score is a list for query-positive and query-negative scores.

Note that `"positive_id"` is not necessarily a positive, and `"negative_id"` is not necessarily a negative. The scores indicate the similarity.

### `list` subset

* Columns: "query_id", "corpus_id", "score"
* Column types: `int`, `list[int]`, `list[float]`
* Examples:
```python
{
"query_id": 571018,
"corpus_id": [7349777, 6948601, 5129919, 6717931, 1065943, 1626276, 981824, 6449111, 1028927, 2524942, 5810175, 6236527, 7179545, 168979, 150383, 168983, 7027047, 3559703, 8768336, 5476579, 915244, 2202253, 1743842, 7727041, 1036624, 8432142, 2236979, 724018, 7179544, 7349780, 7179539, 6072080, 7790852, 4873670, 4389296, 2305477, 1626275, 291845, 1743847, 1508485, 4298457, 1831337, 1760417, 8768340, 8432143, 1971355, 1133925, 2105819, 168975, 5132446, 1316646, 1065945, 7349776, 6717930, 2305472, 8768339, 8768341, 6717927, 7179547, 7491026, 4903324, 1516443, 1065951, 6717926, 4779313, 1381778, 7349774, 6717928, 7349778, 1692036, 168976, 7004464, 5129916, 6243357, 1682970, 2174051, 2735530, 7097201, 4316878, 1508484, 1951254, 2740235, 7790853, 6893978, 4816777, 1191538, 7027046, 165888, 7027044, 1833474, 1065944, 7027050, 7790847, 6717925, 3911285, 7862900, 1065947, 1279944, 8818003, 2174049, 7179546, 978303, 1629126, 4359059, 1891131, 7032037, 8674123, 269779, 371192, 5423524, 150253, 8768342, 8567249, 1833477, 22448, 7862904, 4298455, 7448360, 8768334, 8417201, 2305474, 1283525, 4377211, 7790851, 6243359, 1065948, 7491025, 5437, 1891129, 168974, 7491029, 5129920, 6717924, 468898, 1065952, 2305471, 4903323, 4316880, 8768338, 8184295, 1065946],
"score": [10.257501602172852, 3.5812907218933105, 8.257364273071289, 8.866464614868164, 5.258519172668457, 4.193713188171387, 8.563857078552246, 4.907355785369873, 7.617893695831299, 1.5268436670303345, -0.6152520179748535, -2.9456772804260254, 10.018341064453125, 10.202350616455078, 1.7948371171951294, 8.693719863891602, 4.469407081604004, 0.4720204472541809, -1.00309157371521, 1.8172178268432617, 1.7467658519744873, -1.4857474565505981, 4.076294422149658, -1.777407169342041, -0.7370984554290771, 4.278080463409424, -1.0950472354888916, 2.5531094074249268, 10.004817008972168, 10.176589965820312, 8.594615936279297, 1.8897120952606201, 7.299615859985352, 8.61693000793457, 0.10016749799251556, 6.883630752563477, 10.320749282836914, 0.7852426171302795, 2.7261080741882324, 2.0838329792022705, 1.8327460289001465, -0.6380012035369873, 0.926922082901001, 4.037473201751709, 2.498434543609619, 1.148393154144287, 0.13919004797935486, 3.2860398292541504, 10.097441673278809, 2.575753688812256, 1.7576978206634521, 10.210726737976074, 9.687068939208984, 8.633060455322266, 7.698808193206787, 8.400606155395508, 6.934174060821533, 6.636131286621094, 7.153725624084473, 7.550482273101807, 6.602751731872559, 5.704792022705078, 5.986926555633545, 4.501513481140137, 5.526628017425537, 4.542888164520264, 7.835060119628906, 6.569742202758789, 4.466667175292969, -0.17466464638710022, 5.990896701812744, 4.383068084716797, 5.085425853729248, 6.489709854125977, 3.4293251037597656, 4.946746826171875, 5.910137176513672, 5.161900520324707, 1.4832103252410889, 4.817190170288086, 3.958622694015503, 1.8736721277236938, 6.366949081420898, 4.05584716796875, 4.808823585510254, 2.6205739974975586, 3.1121416091918945, 4.710823059082031, 3.8835949897766113, 7.4977803230285645, 7.494195938110352, 7.061235427856445, 7.726348876953125, 5.88195276260376, 4.692541122436523, 3.5332908630371094, 4.462759017944336, -4.2239532470703125, -2.5660009384155273, 6.035939693450928, 5.124550819396973, 7.7053680419921875, -2.0111143589019775, -0.8396145105361938, 7.110054969787598, 3.5912249088287354, -0.17257803678512573, 0.5216665267944336, 3.553079128265381, -5.091264724731445, 3.0037851333618164, 6.739882469177246, 1.2511098384857178, 5.766049385070801, -4.700324058532715, 2.5425989627838135, 1.9228131771087646, -1.8639280796051025, 3.136643886566162, 0.423944354057312, 2.2642807960510254, 3.421323299407959, -3.7783584594726562, 8.579450607299805, 9.007328987121582, 6.923714637756348, 6.49301290512085, 6.645390033721924, 3.557553291320801, 6.487471580505371, 4.421983242034912, 4.1287055015563965, 6.218915939331055, 6.673498153686523, 4.962984085083008, 4.784761428833008, 3.790182590484619, 3.781992197036743, 5.345108509063721, 4.898831367492676, 4.420433044433594],
}
```
* Collection Strategy: Mining negatives using an embedding model, and rescoring the query/answer pairs. This is all data, grouped per query_id.

搜集汇总

背景与挑战

背景概述

该数据集基于MS MARCO大规模信息检索语料库，包含1.6亿个由cross-encoder/ms-marco-MiniLM-L6-v2模型生成的交叉编码器评分。这些未处理的logits评分可用于通过蒸馏方法微调搜索模型。