update model card

README.md

@@ -1,88 +1,95 @@
----
-pipeline_tag: sentence-similarity
-tags:
-  - feature-extraction
-  - sentence-similarity
-language:
-  - en
----
-
-# Model Card for `vectorizer-v1-S-en`
-
-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-v1-S-en`
-
-## Supported Languages
-
-The model was trained and tested in the following languages:
-
-- English
-
-## Scores
-
-| Metric                 | Value |
-|:-----------------------|------:|
-| Relevance (Recall@100) | 0.456 |
-
-Note that the relevance score is computed as an average over 14 retrieval datasets (see
-[details below](#evaluation-metrics)).
-
-## Inference Times
-
-| GPU        | Batch size 1 (at query time) | Batch size 32 (at indexing) |
-|:-----------|-----------------------------:|----------------------------:|
-| NVIDIA A10 |                         2 ms |                       14 ms |
-| NVIDIA T4  |                         4 ms |                       52 ms |
-
-The inference times only measure the time the model takes to process a single batch, it does not include pre- or
-post-processing steps like the tokenization.
-
-## Requirements
-
-- Minimal Sinequa version: 11.10.0
-- GPU memory usage: 330 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.
-
-## Model Details
-
-### Overview
-
-- Number of parameters: 29 million
-- Base language model: [English BERT-Small](https://huggingface.co/google/bert_uncased_L-4_H-512_A-8)
-- Insensitive to casing and accents
-- Output dimensions: 256 (reduced with an additional dense layer)
-- Training procedure: A first model was trained with query-passage pairs, using the in-batch negative strategy with [this loss](https://www.sbert.net/docs/package_reference/losses.html#multiplenegativesrankingloss). A second model was then trained on query-passage-negative triplets with negatives mined from the previous model, like a variant of [ANCE](https://arxiv.org/pdf/2007.00808.pdf) but with different hyper parameters.
-
-### Training Data
-
-The model was trained on a Sinequa curated version of Google's [Natural Questions](https://ai.google.com/research/NaturalQuestions).
-
-### Evaluation Metrics
-
-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           |      0.456 |
-|                   |            |
-| Arguana           |      0.832 |
-| CLIMATE-FEVER     |      0.342 |
-| DBPedia Entity    |      0.299 |
-| FEVER             |      0.660 |
-| FiQA-2018         |      0.301 |
-| HotpotQA          |      0.434 |
-| MS MARCO          |      0.610 |
-| NFCorpus          |      0.159 |
-| NQ                |      0.671 |
-| Quora             |      0.966 |
-| SCIDOCS           |      0.194 |
-| SciFact           |      0.592 |
-| TREC-COVID        |      0.037 |
-| Webis-Touche-2020 |      0.285 |
+---
+pipeline_tag: sentence-similarity
+tags:
+  - feature-extraction
+  - sentence-similarity
+language:
+  - en
+---
+
+# Model Card for `vectorizer-v1-S-en`
+
+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-v1-S-en`
+
+## Supported Languages
+
+The model was trained and tested in the following languages:
+
+- English
+
+## Scores
+
+| Metric                 | Value |
+|:-----------------------|------:|
+| Relevance (Recall@100) | 0.456 |
+
+Note that the relevance score is computed as an average over 14 retrieval datasets (see
+[details below](#evaluation-metrics)).
+
+## Inference Times
+
+| GPU                                       | Quantization type |  Batch size 1  |  Batch size 32 |
+|:------------------------------------------|:------------------|---------------:|---------------:|
+| NVIDIA A10                                | FP16              |           1 ms |           4 ms |
+| NVIDIA A10                                | FP32              |           2 ms |          13 ms |
+| NVIDIA T4                                 | FP16              |           1 ms |          13 ms |
+| NVIDIA T4                                 | FP32              |           2 ms |          52 ms |
+| NVIDIA L4                                 | FP16              |           1 ms |           5 ms |
+| NVIDIA L4                                 | FP32              |           2 ms |          18 ms |
+
+## Gpu Memory usage
+
+| Quantization type                                |   Memory   |
+|:-------------------------------------------------|-----------:|
+| FP16                                             |    300 MiB |
+| FP32                                             |    500 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: 29 million
+- Base language model: [English BERT-Small](https://huggingface.co/google/bert_uncased_L-4_H-512_A-8)
+- Insensitive to casing and accents
+- Output dimensions: 256 (reduced with an additional dense layer)
+- Training procedure: A first model was trained with query-passage pairs, using the in-batch negative strategy with [this loss](https://www.sbert.net/docs/package_reference/losses.html#multiplenegativesrankingloss). A second model was then trained on query-passage-negative triplets with negatives mined from the previous model, like a variant of [ANCE](https://arxiv.org/pdf/2007.00808.pdf) but with different hyper parameters.
+
+### Training Data
+
+The model was trained on a Sinequa curated version of Google's [Natural Questions](https://ai.google.com/research/NaturalQuestions).
+
+### Evaluation Metrics
+
+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           |      0.456 |
+|                   |            |
+| Arguana           |      0.832 |
+| CLIMATE-FEVER     |      0.342 |
+| DBPedia Entity    |      0.299 |
+| FEVER             |      0.660 |
+| FiQA-2018         |      0.301 |
+| HotpotQA          |      0.434 |
+| MS MARCO          |      0.610 |
+| NFCorpus          |      0.159 |
+| NQ                |      0.671 |
+| Quora             |      0.966 |
+| SCIDOCS           |      0.194 |
+| SciFact           |      0.592 |
+| TREC-COVID        |      0.037 |
+| Webis-Touche-2020 |      0.285 |