Spaces:

Intel
/

qa_sparse_bert

Runtime error

App Files Files Community

Benjamin Consolvo commited on Jan 5, 2023

Commit

cce3268

1 Parent(s): 3ecf378

no picture

Browse files

Files changed (1) hide show

app.py +0 -2

app.py CHANGED Viewed

@@ -36,8 +36,6 @@ def predict(context,question):
 md = """
 If you came looking for chatGPT, sorry to disappoint, but this is different. This prediction model is designed to answer a question about a text. It is designed to do reading comprehension. The model does not just answer questions in general -- it only works from the text that you provide. However, accomplishing accurate reading comprehension can be a valuable task.
-![](https://images.pexels.com/photos/1906437/pexels-photo-1906437.jpeg)
 The model is based on the Zafrir et al. (2021) paper: [Prune Once for All: Sparse Pre-Trained Language Models](https://arxiv.org/abs/2111.05754). The model can be found here: https://huggingface.co/Intel/bert-base-uncased-squadv1.1-sparse-80-1x4-block-pruneofa. The main idea of this BERT-Base model is that it is much more fast and efficient in deployment than its dense counterpart: (https://huggingface.co/csarron/bert-base-uncased-squad-v1). It has had weight pruning and model distillation applied to create a sparse weight pattern that is maintained even after fine-tuning has been applied. According to Zafrir et al. (2016), their "results show the best compression-to-accuracy ratio for BERT-Base". This model is still in FP32, but can be quantized to INT8 with the [Intel® Neural Compressor](https://github.com/intel/neural-compressor) for further compression.
 The training dataset used is the English Wikipedia dataset (2500M words), and then fine-tuned on the SQuADv1.1 dataset containing 89K training examples by Rajpurkar et al. (2016): [100, 000+ Questions for Machine Comprehension of Text](https://arxiv.org/abs/1606.05250).

 md = """
 If you came looking for chatGPT, sorry to disappoint, but this is different. This prediction model is designed to answer a question about a text. It is designed to do reading comprehension. The model does not just answer questions in general -- it only works from the text that you provide. However, accomplishing accurate reading comprehension can be a valuable task.
 The model is based on the Zafrir et al. (2021) paper: [Prune Once for All: Sparse Pre-Trained Language Models](https://arxiv.org/abs/2111.05754). The model can be found here: https://huggingface.co/Intel/bert-base-uncased-squadv1.1-sparse-80-1x4-block-pruneofa. The main idea of this BERT-Base model is that it is much more fast and efficient in deployment than its dense counterpart: (https://huggingface.co/csarron/bert-base-uncased-squad-v1). It has had weight pruning and model distillation applied to create a sparse weight pattern that is maintained even after fine-tuning has been applied. According to Zafrir et al. (2016), their "results show the best compression-to-accuracy ratio for BERT-Base". This model is still in FP32, but can be quantized to INT8 with the [Intel® Neural Compressor](https://github.com/intel/neural-compressor) for further compression.
 The training dataset used is the English Wikipedia dataset (2500M words), and then fine-tuned on the SQuADv1.1 dataset containing 89K training examples by Rajpurkar et al. (2016): [100, 000+ Questions for Machine Comprehension of Text](https://arxiv.org/abs/1606.05250).