|
---
|
|
language:
|
|
- de
|
|
- en
|
|
- es
|
|
- fr
|
|
- it
|
|
- ja
|
|
- nl
|
|
- pl
|
|
- pt
|
|
- zh
|
|
---
|
|
|
|
# Model Card for `passage-ranker.pistachio`
|
|
|
|
This model is a passage ranker developed by Sinequa. It produces a relevance score given a query-passage pair and is used to order search results.
|
|
|
|
Model name: `passage-ranker.pistachio`
|
|
|
|
## Supported Languages
|
|
|
|
The model was trained and tested in the following languages:
|
|
|
|
- Chinese (simplified)
|
|
- Dutch
|
|
- English
|
|
- French
|
|
- German
|
|
- Italian
|
|
- Japanese
|
|
- Polish
|
|
- Portuguese
|
|
- Spanish
|
|
|
|
Besides the aforementioned languages, basic support can be expected for additional 93 languages that were used during the pretraining of the base model (see
|
|
[list of languages](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages)).
|
|
|
|
## Scores
|
|
|
|
| Metric | Value |
|
|
|:----------------------------|------:|
|
|
| English Relevance (NDCG@10) | 0.474 |
|
|
| Polish Relevance (NDCG@10) | 0.380 |
|
|
|
|
Note that the relevance score is 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 | 2 ms | 28 ms |
|
|
| NVIDIA A10 | FP32 | 4 ms | 82 ms |
|
|
| NVIDIA T4 | FP16 | 3 ms | 65 ms |
|
|
| NVIDIA T4 | FP32 | 14 ms | 369 ms |
|
|
| NVIDIA L4 | FP16 | 3 ms | 38 ms |
|
|
| NVIDIA L4 | FP32 | 5 ms | 123 ms |
|
|
|
|
## Gpu Memory usage
|
|
|
|
| Quantization type | Memory |
|
|
|:-------------------------------------------------|-----------:|
|
|
| FP16 | 850 MiB |
|
|
| FP32 | 1200 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: 167 million
|
|
- Base language model: [Multilingual BERT-Base](https://huggingface.co/bert-base-multilingual-uncased)
|
|
- Insensitive to casing and accents
|
|
- Training procedure: [MonoBERT](https://arxiv.org/abs/1901.04085)
|
|
|
|
### Training Data
|
|
|
|
- MS MARCO Passage Ranking
|
|
([Paper](https://arxiv.org/abs/1611.09268),
|
|
[Official Page](https://microsoft.github.io/msmarco/),
|
|
[English & translated datasets on the HF dataset hub](https://huggingface.co/datasets/unicamp-dl/mmarco), [translated dataset in Polish on the HF dataset hub](https://huggingface.co/datasets/clarin-knext/msmarco-pl))
|
|
- Original English dataset
|
|
- Translated datasets for the other nine supported languages
|
|
|
|
### 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 | NDCG@10 |
|
|
|:------------------|--------:|
|
|
| Average | 0.474 |
|
|
| | |
|
|
| Arguana | 0.539 |
|
|
| CLIMATE-FEVER | 0.230 |
|
|
| DBPedia Entity | 0.369 |
|
|
| FEVER | 0.765 |
|
|
| FiQA-2018 | 0.329 |
|
|
| HotpotQA | 0.694 |
|
|
| MS MARCO | 0.413 |
|
|
| NFCorpus | 0.337 |
|
|
| NQ | 0.486 |
|
|
| Quora | 0.714 |
|
|
| SCIDOCS | 0.144 |
|
|
| SciFact | 0.649 |
|
|
| TREC-COVID | 0.651 |
|
|
| Webis-Touche-2020 | 0.312 |
|
|
|
|
#### Polish
|
|
|
|
This model has polish capacities, that are being evaluated over a subset of
|
|
the [PIRBenchmark](https://github.com/sdadas/pirb) with BM25 as the first stage retrieval.
|
|
|
|
|
|
| Dataset | NDCG@10 |
|
|
|:--------------|--------:|
|
|
| Average | 0.380 |
|
|
| | |
|
|
| arguana-pl | 0.285 |
|
|
| dbpedia-pl | 0.283 |
|
|
| fiqa-pl | 0.223 |
|
|
| hotpotqa-pl | 0.603 |
|
|
| msmarco-pl | 0.259 |
|
|
| nfcorpus-pl | 0.293 |
|
|
| nq-pl | 0.355 |
|
|
| quora-pl | 0.613 |
|
|
| scidocs-pl | 0.128 |
|
|
| scifact-pl | 0.581 |
|
|
| trec-covid-pl | 0.560 |
|
|
|
|
#### 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 | NDCG@10 |
|
|
|:----------------------|--------:|
|
|
| Chinese (simplified) | 0.454 |
|
|
| French | 0.439 |
|
|
| German | 0.418 |
|
|
| Japanese | 0.517 |
|
|
| Spanish | 0.487 |
|
|
|
