File size: 5,902 Bytes
a538308
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6193cfd
a538308
 
 
 
 
 
 
dc6d49c
a538308
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
---
pipeline_tag: sentence-similarity
tags:
  - feature-extraction
  - sentence-similarity
language:
  - de
  - en
  - es
  - fr
  - it
  - nl
  - ja
  - pt
  - zh
  - pl
---

# Model Card for `vectorizer.guava`

This model is a vectorizer developed by Sinequa. It produces an embedding vector given a passage or a query. The
passage vectors are stored in our vector index and the query vector is used at query time to look up relevant passages
in the index.

Model name: `vectorizer.guava`

## Supported Languages

The model was trained and tested in the following languages:

- English
- French
- German
- Spanish
- Italian
- Dutch
- Japanese
- Portuguese
- Chinese (simplified)
- Chinese (traditional)
- Polish

Besides these languages, basic support can be expected for additional 91 languages that were used during the pretraining
of the base model (see Appendix A of XLM-R paper).

## Scores

| Metric                         | Value |
|:-------------------------------|------:|
| English Relevance (Recall@100) | 0.616 |

Note that the relevance scores are computed as an average over several retrieval datasets (see
[details below](#evaluation-metrics)).

## Inference Times

| GPU                                       | Quantization type |  Batch size 1  |  Batch size 32 |
|:------------------------------------------|:------------------|---------------:|---------------:|
| NVIDIA A10                                | FP16              |           1 ms |           5 ms |
| NVIDIA A10                                | FP32              |           2 ms |          18 ms |
| NVIDIA T4                                 | FP16              |           1 ms |          12 ms |
| NVIDIA T4                                 | FP32              |           3 ms |          52 ms |
| NVIDIA L4                                 | FP16              |           2 ms |           5 ms |
| NVIDIA L4                                 | FP32              |           4 ms |          24 ms |

## Gpu Memory usage

| Quantization type                                |   Memory   |
|:-------------------------------------------------|-----------:|
| FP16                                             |    550 MiB |
| FP32                                             |   1050 MiB |

Note that GPU memory usage only includes how much GPU memory the actual model consumes on an NVIDIA T4 GPU with a batch
size of 32. It does not include the fix amount of memory that is consumed by the ONNX Runtime upon initialization which
can be around 0.5 to 1 GiB depending on the used GPU.

## Requirements

- Minimal Sinequa version: 11.10.0
- Minimal Sinequa version for using FP16 models and GPUs with CUDA compute capability of 8.9+ (like NVIDIA L4): 11.11.0
- [Cuda compute capability](https://developer.nvidia.com/cuda-gpus): above 5.0 (above 6.0 for FP16 use)

## Model Details

### Overview

- Number of parameters: 107 million
- Base language
  model: [mMiniLMv2-L6-H384-distilled-from-XLMR-Large](https://huggingface.co/nreimers/mMiniLMv2-L6-H384-distilled-from-XLMR-Large) ([Paper](https://arxiv.org/abs/2012.15828), [GitHub](https://github.com/microsoft/unilm/tree/master/minilm))
- Insensitive to casing and accents
- Output dimensions: 256 (reduced with an additional dense layer)
- Training procedure: Query-passage-negative triplets for datasets that have mined hard negative data, Query-passage
  pairs for the rest. Number of negatives is augmented with in-batch negative strategy

### Training Data

The model have been trained using all datasets that are cited in
the [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) model.
In addition to that, this model has been trained on the datasets cited
in [this paper](https://arxiv.org/pdf/2108.13897.pdf) on the first 9 aforementioned languages.
It has also been trained on [this dataset](https://huggingface.co/datasets/clarin-knext/msmarco-pl) for polish capacities, and a translated version of msmarco-zh for traditional chinese capacities.

### Evaluation Metrics

#### English

To determine the relevance score, we averaged the results that we obtained when evaluating on the datasets of the
[BEIR benchmark](https://github.com/beir-cellar/beir). Note that all these datasets are in **English**.

| Dataset           | Recall@100 |
|:------------------|-----------:|
| Average           |      0.616 |
|                   |            |
| Arguana           |      0.956 |
| CLIMATE-FEVER     |      0.471 |
| DBPedia Entity    |      0.379 |
| FEVER             |      0.824 |
| FiQA-2018         |      0.642 |
| HotpotQA          |      0.579 |
| MS MARCO          |      0.85 |
| NFCorpus          |      0.289 |
| NQ                |      0.765 |
| Quora             |      0.993 |
| SCIDOCS           |      0.467 |
| SciFact           |      0.899 |
| TREC-COVID        |      0.104 |
| Webis-Touche-2020 |      0.407 |

#### Traditional Chinese

This model has traditional chinese capacities, that are being evaluated over the same dev set at msmarco-zh, translated in traditional chinese.

| Dataset                          | Recall@100 |
|:---------------------------------|-----------:|
| msmarco-zh-traditional           |      0.738 |

In comparison, [raspberry](https://huggingface.co/sinequa/vectorizer.raspberry) scores a 0.693 on this dataset.


#### Other languages

We evaluated the model on the datasets of the [MIRACL benchmark](https://github.com/project-miracl/miracl) to test its
multilingual capacities. Note that not all training languages are part of the benchmark, so we only report the metrics
for the existing languages.

| Language              | Recall@100 |
|:----------------------|-----------:|
| French                |      0.672 |
| German                |      0.594 |
| Spanish               |      0.632 |
| Japanese              |      0.603 |
| Chinese (simplified)  |      0.702 |