|
--- |
|
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 | |
|
|
