Model Card for answer-finder.yuzu

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.

Model name: answer-finder.yuzu

Supported Languages

The model was trained and tested in the following languages:

• Japanese

Besides the aforementioned languages, basic support can be expected for the 104 languages that were used during the pretraining of the base model (See original repository).

Scores

Metric	Value
F1 Score on JSQuAD with Hugging Face evaluation pipeline	92.1
F1 Score on JSQuAD with Haystack evaluation pipeline	91.5

Inference Time

GPU	Quantization type	Batch size 1	Batch size 32
NVIDIA A10	FP16	17 ms	27 ms
NVIDIA A10	FP32	4 ms	88 ms
NVIDIA T4	FP16	3 ms	64 ms
NVIDIA T4	FP32	15 ms	374 ms
NVIDIA L4	FP16	3 ms	39 ms
NVIDIA L4	FP32	5 ms	125 ms

Note that the Answer Finder models are only used at query time.

Gpu Memory usage

Quantization type	Memory
FP16	950 MiB
FP32	1350 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: 110 million
• Base language model: bert-base-multilingual-cased
• Sensitive to casing and accents

Training Data

• JSQuAD see Paper
• Japanese translation of SQuAD v2 "impossible" query-passage pairs