Correct dataset definition
Browse files
README.md
CHANGED
|
@@ -1,104 +1,103 @@
|
|
| 1 |
-
---
|
| 2 |
-
language:
|
| 3 |
-
- de
|
| 4 |
-
- en
|
| 5 |
-
- es
|
| 6 |
-
- fr
|
| 7 |
-
---
|
| 8 |
-
|
| 9 |
-
# Model Card for `passage-ranker-v1-XS-multilingual`
|
| 10 |
-
|
| 11 |
-
This model is a passage ranker developed by Sinequa. It produces a relevance score given a query-passage pair and is
|
| 12 |
-
used to order search results.
|
| 13 |
-
|
| 14 |
-
Model name: `passage-ranker-v1-XS-multilingual`
|
| 15 |
-
|
| 16 |
-
## Supported Languages
|
| 17 |
-
|
| 18 |
-
The model was trained and tested in the following languages:
|
| 19 |
-
|
| 20 |
-
- English
|
| 21 |
-
- French
|
| 22 |
-
- German
|
| 23 |
-
- Spanish
|
| 24 |
-
|
| 25 |
-
## Scores
|
| 26 |
-
|
| 27 |
-
| Metric | Value |
|
| 28 |
-
|:--------------------|------:|
|
| 29 |
-
| Relevance (NDCG@10) | 0.453 |
|
| 30 |
-
|
| 31 |
-
Note that the relevance score is computed as an average over 14 retrieval datasets (see
|
| 32 |
-
[details below](#evaluation-metrics)).
|
| 33 |
-
|
| 34 |
-
## Inference Times
|
| 35 |
-
|
| 36 |
-
| GPU | Batch size 32 |
|
| 37 |
-
|:-----------|--------------:|
|
| 38 |
-
| NVIDIA A10 | 8 ms |
|
| 39 |
-
| NVIDIA T4 | 21 ms |
|
| 40 |
-
|
| 41 |
-
The inference times only measure the time the model takes to process a single batch, it does not include pre- or
|
| 42 |
-
post-processing steps like the tokenization.
|
| 43 |
-
|
| 44 |
-
## Requirements
|
| 45 |
-
|
| 46 |
-
- Minimal Sinequa version: 11.10.0
|
| 47 |
-
- GPU memory usage: 300 MiB
|
| 48 |
-
|
| 49 |
-
Note that GPU memory usage only includes how much GPU memory the actual model consumes on an NVIDIA T4 GPU with a batch
|
| 50 |
-
size of 32. It does not include the fix amount of memory that is consumed by the ONNX Runtime upon initialization which
|
| 51 |
-
can be around 0.5 to 1 GiB depending on the used GPU.
|
| 52 |
-
|
| 53 |
-
## Model Details
|
| 54 |
-
|
| 55 |
-
### Overview
|
| 56 |
-
|
| 57 |
-
- Number of parameters: 16 million
|
| 58 |
-
- Base language model: Homegrown Sinequa BERT-Mini ([Paper](https://arxiv.org/abs/1908.08962)) pretrained in the four
|
| 59 |
-
supported languages
|
| 60 |
-
- Insensitive to casing and accents
|
| 61 |
-
- Training procedure: [MonoBERT](https://arxiv.org/abs/1901.04085)
|
| 62 |
-
|
| 63 |
-
### Training Data
|
| 64 |
-
|
| 65 |
-
-
|
| 66 |
-
([Paper](https://arxiv.org/abs/
|
| 67 |
-
[Official Page](https://
|
| 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 |
-
| Spanish | 0.416 |
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- de
|
| 4 |
+
- en
|
| 5 |
+
- es
|
| 6 |
+
- fr
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# Model Card for `passage-ranker-v1-XS-multilingual`
|
| 10 |
+
|
| 11 |
+
This model is a passage ranker developed by Sinequa. It produces a relevance score given a query-passage pair and is
|
| 12 |
+
used to order search results.
|
| 13 |
+
|
| 14 |
+
Model name: `passage-ranker-v1-XS-multilingual`
|
| 15 |
+
|
| 16 |
+
## Supported Languages
|
| 17 |
+
|
| 18 |
+
The model was trained and tested in the following languages:
|
| 19 |
+
|
| 20 |
+
- English
|
| 21 |
+
- French
|
| 22 |
+
- German
|
| 23 |
+
- Spanish
|
| 24 |
+
|
| 25 |
+
## Scores
|
| 26 |
+
|
| 27 |
+
| Metric | Value |
|
| 28 |
+
|:--------------------|------:|
|
| 29 |
+
| Relevance (NDCG@10) | 0.453 |
|
| 30 |
+
|
| 31 |
+
Note that the relevance score is computed as an average over 14 retrieval datasets (see
|
| 32 |
+
[details below](#evaluation-metrics)).
|
| 33 |
+
|
| 34 |
+
## Inference Times
|
| 35 |
+
|
| 36 |
+
| GPU | Batch size 32 |
|
| 37 |
+
|:-----------|--------------:|
|
| 38 |
+
| NVIDIA A10 | 8 ms |
|
| 39 |
+
| NVIDIA T4 | 21 ms |
|
| 40 |
+
|
| 41 |
+
The inference times only measure the time the model takes to process a single batch, it does not include pre- or
|
| 42 |
+
post-processing steps like the tokenization.
|
| 43 |
+
|
| 44 |
+
## Requirements
|
| 45 |
+
|
| 46 |
+
- Minimal Sinequa version: 11.10.0
|
| 47 |
+
- GPU memory usage: 300 MiB
|
| 48 |
+
|
| 49 |
+
Note that GPU memory usage only includes how much GPU memory the actual model consumes on an NVIDIA T4 GPU with a batch
|
| 50 |
+
size of 32. It does not include the fix amount of memory that is consumed by the ONNX Runtime upon initialization which
|
| 51 |
+
can be around 0.5 to 1 GiB depending on the used GPU.
|
| 52 |
+
|
| 53 |
+
## Model Details
|
| 54 |
+
|
| 55 |
+
### Overview
|
| 56 |
+
|
| 57 |
+
- Number of parameters: 16 million
|
| 58 |
+
- Base language model: Homegrown Sinequa BERT-Mini ([Paper](https://arxiv.org/abs/1908.08962)) pretrained in the four
|
| 59 |
+
supported languages
|
| 60 |
+
- Insensitive to casing and accents
|
| 61 |
+
- Training procedure: [MonoBERT](https://arxiv.org/abs/1901.04085)
|
| 62 |
+
|
| 63 |
+
### Training Data
|
| 64 |
+
|
| 65 |
+
- Probably-Asked Questions
|
| 66 |
+
([Paper](https://arxiv.org/abs/2102.07033),
|
| 67 |
+
[Official Page](https://github.com/facebookresearch/PAQ))
|
| 68 |
+
- Original English dataset
|
| 69 |
+
- Translated datasets for the other three supported languages
|
| 70 |
+
|
| 71 |
+
### Evaluation Metrics
|
| 72 |
+
|
| 73 |
+
To determine the relevance score, we averaged the results that we obtained when evaluating on the datasets of the
|
| 74 |
+
[BEIR benchmark](https://github.com/beir-cellar/beir). Note that all these datasets are in English.
|
| 75 |
+
|
| 76 |
+
| Dataset | NDCG@10 |
|
| 77 |
+
|:------------------|--------:|
|
| 78 |
+
| Average | 0.453 |
|
| 79 |
+
| | |
|
| 80 |
+
| Arguana | 0.516 |
|
| 81 |
+
| CLIMATE-FEVER | 0.159 |
|
| 82 |
+
| DBPedia Entity | 0.355 |
|
| 83 |
+
| FEVER | 0.729 |
|
| 84 |
+
| FiQA-2018 | 0.282 |
|
| 85 |
+
| HotpotQA | 0.688 |
|
| 86 |
+
| MS MARCO | 0.334 |
|
| 87 |
+
| NFCorpus | 0.341 |
|
| 88 |
+
| NQ | 0.438 |
|
| 89 |
+
| Quora | 0.726 |
|
| 90 |
+
| SCIDOCS | 0.143 |
|
| 91 |
+
| SciFact | 0.630 |
|
| 92 |
+
| TREC-COVID | 0.664 |
|
| 93 |
+
| Webis-Touche-2020 | 0.337 |
|
| 94 |
+
|
| 95 |
+
We evaluated the model on the datasets of the [MIRACL benchmark](https://github.com/project-miracl/miracl) to test its
|
| 96 |
+
multilingual capacities. Note that not all training languages are part of the benchmark, so we only report the metrics
|
| 97 |
+
for the existing languages.
|
| 98 |
+
|
| 99 |
+
| Language | NDCG@10 |
|
| 100 |
+
|:---------|--------:|
|
| 101 |
+
| French | 0.346 |
|
| 102 |
+
| German | 0.368 |
|
| 103 |
+
| Spanish | 0.416 |
|
|
|