英文

模型详情:INT8 DistilBERT base uncased finetuned SST-2

该模型是一个针对情感分类的DistilBERT模型进行微调的结果,在原始FP32模型上进行了INT8量化(训练后静态量化)。同一个模型以不同格式提供:PyTorch和ONNX。

Model Detail Description
Model Authors - Company Intel
Date March 29, 2022 for PyTorch model & February 3, 2023 for ONNX model
Version 1
Type NLP DistilBERT (INT8) - Sentiment Classification (+/-)
Paper or Other Resources 1235321
License Apache 2.0
Questions or Comments 1236321 and 1237321
Intended Use Description
Primary intended uses Inference for sentiment classification (classifying whether a statement is positive or negative)
Primary intended users Anyone
Out-of-scope uses This model is already fine-tuned and quantized to INT8. It is not suitable for further fine-tuning in this form. To fine-tune your own model, you can start with 1238321 . The model should not be used to intentionally create hostile or alienating environments for people.
使用最佳英特尔进行PyTorch模型加载
from optimum.intel.neural_compressor import INCModelForSequenceClassification

model_id = "Intel/distilbert-base-uncased-finetuned-sst-2-english-int8-static"
int8_model = INCModelForSequenceClassification.from_pretrained(model_id)
使用最佳英特尔进行ONNX模型加载:
from optimum.onnxruntime import ORTModelForSequenceClassification

model_id = "Intel/distilbert-base-uncased-finetuned-sst-2-english-int8-static"
int8_model = ORTModelForSequenceClassification.from_pretrained(model_id)
Factors Description
Groups Movie reviewers from the internet
Instrumentation Text movie single-sentence reviews taken from 4 authors. More information can be found in the original paper by 1239321
Environment -
Card Prompts Model deployment on alternate hardware and software can change model performance
Metrics Description
Model performance measures Accuracy
Decision thresholds -
Approaches to uncertainty and variability -
PyTorch INT8 ONNX INT8 FP32
Accuracy (eval-accuracy) 0.9037 0.9071 0.9106
Model Size (MB) 65 89 255
Training and Evaluation Data Description
Datasets The dataset can be found here: 12310321 . There dataset has a total of 215,154 unique phrases, annotated by 3 human judges.
Motivation Dataset was chosen to showcase the benefits of quantization on an NLP classification task with the 12311321 and 12312321
Preprocessing The calibration dataloader is the train dataloader. The default calibration sampling size 100 isn't divisible exactly by batch size 8, so the real sampling size is 104.
Quantitative Analyses Description
Unitary results The model was only evaluated on accuracy. There is no available comparison between evaluation factors.
Intersectional results There is no available comparison between the intersection of evaluated factors.
Ethical Considerations Description
Data The data that make up the model are movie reviews from authors on the internet.
Human life The model is not intended to inform decisions central to human life or flourishing. It is an aggregated set of movie reviews from the internet.
Mitigations No additional risk mitigation strategies were considered during model development.
Risks and harms The data are biased toward the particular reviewers' opinions and the judges (labelers) of the data. Significant research has explored bias and fairness issues with language models (see, e.g., 12313321 , and 12314321 ). Predictions generated by the model may include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups. Beyond this, the extent of the risks involved by using the model remain unknown.
Use cases -
Caveats and Recommendations
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. There are no additional caveats or recommendations for this model.

Bibtex条目和引用信息

@misc{distilbert-base-uncased-finetuned-sst-2-english-int8-static
  author    = {Xin He, Yu Wenz},
  title     = {distilbert-base-uncased-finetuned-sst-2-english-int8-static},
  year      = {2022},
  url       = {https://huggingface.co/Intel/distilbert-base-uncased-finetuned-sst-2-english-int8-static},
}