Model Card for vectorizer.hazelnut

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.hazelnut

Supported Languages

The model was trained and tested in the following languages:

• English
• French
• German
• Spanish
• Italian
• Dutch
• Japanese
• Portuguese
• Chinese (simplified)
• 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)	0.590
Polish Relevance (Recall@100)	0.543

Note that the relevance scores are computed as an average over several retrieval datasets (see details below).

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: 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 (Paper, GitHub)
• 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 model. In addition to that, this model has been trained on the datasets cited in this paper on the first 9 aforementioned languages. It has also been trained on this dataset for polish 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. Note that all these datasets are in English.

Dataset	Recall@100
Average	0.590
Arguana	0.961
CLIMATE-FEVER	0.432
DBPedia Entity	0.371
FEVER	0.723
FiQA-2018	0.611
HotpotQA	0.564
MS MARCO	0.825
NFCorpus	0.266
NQ	0.722
Quora	0.991
SCIDOCS	0.426
SciFact	0.864
TREC-COVID	0.092
Webis-Touche-2020	0.415

Polish

This model has polish capacities, that are being evaluated over a subset of the PIRBenchmark.

Dataset	Recall@100
Average	0.534
arguana-pl	0.909
dbpedia-pl	0.282
fiqa-pl	0.439
hotpotqa-pl	0.530
msmarco-pl	0.694
nfcorpus-pl	0.218
nq-pl	0.697
quora-pl	0.949
scidocs-pl	0.291
scifact-pl	0.805
trec-covid-pl	0.059

Other languages

We evaluated the model on the datasets of the MIRACL benchmark 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	0.649
German	0.598
Spanish	0.609
Japanese	0.623
Chinese (simplified)	0.707