being more general about minimal version with fp16 and gpus cc 8.9+
Browse files
README.md
CHANGED
|
@@ -1,78 +1,87 @@
|
|
| 1 |
-
---
|
| 2 |
-
language:
|
| 3 |
-
- de
|
| 4 |
-
- en
|
| 5 |
-
- es
|
| 6 |
-
- fr
|
| 7 |
-
---
|
| 8 |
-
|
| 9 |
-
# Model Card for `answer-finder-v1-L-multilingual`
|
| 10 |
-
|
| 11 |
-
This model is a question answering model developed by Sinequa. It produces two lists of logit scores corresponding to
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
-
|
| 21 |
-
-
|
| 22 |
-
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
| F1 Score on SQuAD v2 EN with
|
| 30 |
-
| F1 Score on SQuAD v2
|
| 31 |
-
| F1 Score on SQuAD v2
|
| 32 |
-
| F1 Score on SQuAD v2
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
| NVIDIA A10 |
|
| 40 |
-
| NVIDIA
|
| 41 |
-
| NVIDIA T4 |
|
| 42 |
-
| NVIDIA
|
| 43 |
-
| NVIDIA L4 |
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
-
|
| 70 |
-
|
| 71 |
-
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
-
|
| 78 |
-
- [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- de
|
| 4 |
+
- en
|
| 5 |
+
- es
|
| 6 |
+
- fr
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# Model Card for `answer-finder-v1-L-multilingual`
|
| 10 |
+
|
| 11 |
+
This model is a question answering model developed by Sinequa. It produces two lists of logit scores corresponding to the start token and end token of an answer.
|
| 12 |
+
|
| 13 |
+
Model name: `answer-finder-v1-L-multilingual`
|
| 14 |
+
|
| 15 |
+
## Supported Languages
|
| 16 |
+
|
| 17 |
+
The model was trained and tested in the following languages:
|
| 18 |
+
|
| 19 |
+
- English
|
| 20 |
+
- French
|
| 21 |
+
- German
|
| 22 |
+
- Spanish
|
| 23 |
+
|
| 24 |
+
## Scores
|
| 25 |
+
|
| 26 |
+
| Metric | Value |
|
| 27 |
+
|:--------------------------------------------------------------|-------:|
|
| 28 |
+
| F1 Score on SQuAD v2 EN with Hugging Face evaluation pipeline | 75 |
|
| 29 |
+
| F1 Score on SQuAD v2 EN with Haystack evaluation pipeline | 75 |
|
| 30 |
+
| F1 Score on SQuAD v2 FR with Haystack evaluation pipeline | 73.4 |
|
| 31 |
+
| F1 Score on SQuAD v2 DE with Haystack evaluation pipeline | 90.8 |
|
| 32 |
+
| F1 Score on SQuAD v2 ES with Haystack evaluation pipeline | 67.1 |
|
| 33 |
+
|
| 34 |
+
## Inference Time
|
| 35 |
+
|
| 36 |
+
| GPU | Quantization type | Batch size 1 | Batch size 32 |
|
| 37 |
+
|:------------------------------------------|:------------------|---------------:|---------------:|
|
| 38 |
+
| NVIDIA A10 | FP16 | 2 ms | 30 ms |
|
| 39 |
+
| NVIDIA A10 | FP32 | 4 ms | 83 ms |
|
| 40 |
+
| NVIDIA T4 | FP16 | 3 ms | 65 ms |
|
| 41 |
+
| NVIDIA T4 | FP32 | 14 ms | 373 ms |
|
| 42 |
+
| NVIDIA L4 | FP16 | 2 ms | 38 ms |
|
| 43 |
+
| NVIDIA L4 | FP32 | 5 ms | 124 ms |
|
| 44 |
+
|
| 45 |
+
**Note that the Answer Finder models are only used at query time.**
|
| 46 |
+
|
| 47 |
+
## Gpu Memory usage
|
| 48 |
+
|
| 49 |
+
| Quantization type | Memory |
|
| 50 |
+
|:-------------------------------------------------|-----------:|
|
| 51 |
+
| FP16 | 550 MiB |
|
| 52 |
+
| FP32 | 1050 MiB |
|
| 53 |
+
|
| 54 |
+
Note that GPU memory usage only includes how much GPU memory the actual model consumes on an NVIDIA T4 GPU with a batch
|
| 55 |
+
size of 32. It does not include the fix amount of memory that is consumed by the ONNX Runtime upon initialization which
|
| 56 |
+
can be around 0.5 to 1 GiB depending on the used GPU.
|
| 57 |
+
|
| 58 |
+
## GPU Memory usage
|
| 59 |
+
|
| 60 |
+
| Quantization type | Memory |
|
| 61 |
+
|:-------------------------------------------------|-----------:|
|
| 62 |
+
| FP16 | 547 MiB |
|
| 63 |
+
| FP32 | 1060 MiB |
|
| 64 |
+
|
| 65 |
+
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.
|
| 66 |
+
|
| 67 |
+
## Requirements
|
| 68 |
+
|
| 69 |
+
- Minimal Sinequa version: 11.10.0
|
| 70 |
+
- Minimal Sinequa version for using FP16 models and GPUs with CUDA compute capability of 8.9+ (like NVIDIA L4): 11.11.0
|
| 71 |
+
- [Cuda compute capability](https://developer.nvidia.com/cuda-gpus): above 5.0 (above 6.0 for FP16 use)
|
| 72 |
+
|
| 73 |
+
## Model Details
|
| 74 |
+
|
| 75 |
+
### Overview
|
| 76 |
+
|
| 77 |
+
- Number of parameters: 110 million
|
| 78 |
+
- Base language model: [bert-base-multilingual-cased](https://huggingface.co/bert-base-multilingual-cased)
|
| 79 |
+
pre-trained by Sinequa in English, French, German and Spanish
|
| 80 |
+
- Insensitive to casing and accents
|
| 81 |
+
|
| 82 |
+
### Training Data
|
| 83 |
+
|
| 84 |
+
- [SQuAD v2](https://rajpurkar.github.io/SQuAD-explorer/)
|
| 85 |
+
- [French-SQuAD](https://github.com/Alikabbadj/French-SQuAD) + French translation of SQuAD v2 "impossible" query-passage pairs
|
| 86 |
+
- [GermanQuAD](https://www.deepset.ai/germanquad) + German translation of SQuAD v2 "impossible" query-passage pairs
|
| 87 |
+
- [SQuAD-es-v2](https://github.com/ccasimiro88/TranslateAlignRetrieve)
|