1
---
2
language: en
3
tags:
4
- tapas
5
license: apache-2.0
6
datasets:
7
- wikisql
8
---
9
10
# TAPAS medium model fine-tuned on WikiSQL (in a supervised fashion)
11
12
his model has 2 versions which can be used. The default version corresponds to the `tapas_wikisql_sqa_inter_masklm_medium_reset` checkpoint of the [original Github repository](https://github.com/google-research/tapas).
13
This model was pre-trained on MLM and an additional step which the authors call intermediate pre-training, and then fine-tuned in a chain on [SQA](https://www.microsoft.com/en-us/download/details.aspx?id=54253), and [WikiSQL](https://github.com/salesforce/WikiSQL). It uses relative position embeddings (i.e. resetting the position index at every cell of the table).
14
15
The other (non-default) version which can be used is: 
16
- `no_reset`, which corresponds to `tapas_wikisql_sqa_inter_masklm_medium` (intermediate pre-training, absolute position embeddings). 
17
18
Disclaimer: The team releasing TAPAS did not write a model card for this model so this model card has been written by
19
the Hugging Face team and contributors.
20
21
## Model description
22
23
TAPAS is a BERT-like transformers model pretrained on a large corpus of English data from Wikipedia in a self-supervised fashion. 
24
This means it was pretrained on the raw tables and associated texts only, with no humans labelling them in any way (which is why it
25
can use lots of publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
26
was pretrained with two objectives:
27
28
- Masked language modeling (MLM): taking a (flattened) table and associated context, the model randomly masks 15% of the words in 
29
  the input, then runs the entire (partially masked) sequence through the model. The model then has to predict the masked words. 
30
  This is different from traditional recurrent neural networks (RNNs) that usually see the words one after the other, 
31
  or from autoregressive models like GPT which internally mask the future tokens. It allows the model to learn a bidirectional 
32
  representation of a table and associated text.
33
- Intermediate pre-training: to encourage numerical reasoning on tables, the authors additionally pre-trained the model by creating 
34
  a balanced dataset of millions of syntactically created training examples. Here, the model must predict (classify) whether a sentence 
35
  is supported or refuted by the contents of a table. The training examples are created based on synthetic as well as counterfactual statements.
36
37
This way, the model learns an inner representation of the English language used in tables and associated texts, which can then be used 
38
to extract features useful for downstream tasks such as answering questions about a table, or determining whether a sentence is entailed
39
or refuted by the contents of a table. Fine-tuning is done by adding a cell selection head and aggregation head on top of the pre-trained model, and then jointly train these randomly initialized classification heads with the base model on SQA and WikiSQL. 
40
41
42
## Intended uses & limitations
43
44
You can use this model for answering questions related to a table.
45
46
For code examples, we refer to the documentation of TAPAS on the HuggingFace website. 
47
48
49
## Training procedure
50
51
### Preprocessing
52
53
The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
54
then of the form:
55
56
```
57
[CLS] Question [SEP] Flattened table [SEP]
58
```
59
60
The authors did first convert the WikiSQL dataset into the format of SQA using automatic conversion scripts.
61
62
### Fine-tuning
63
64
The model was fine-tuned on 32 Cloud TPU v3 cores for 50,000 steps with maximum sequence length 512 and batch size of 512.
65
In this setup, fine-tuning takes around 10 hours. The optimizer used is Adam with a learning rate of 6.17164e-5, and a warmup 
66
ratio of 0.1424. See the [paper](https://arxiv.org/abs/2004.02349) for more details (tables 11 and 12). 
67
68
69
### BibTeX entry and citation info
70
71
```bibtex
72
@misc{herzig2020tapas,
73
      title={TAPAS: Weakly Supervised Table Parsing via Pre-training}, 
74
      author={Jonathan Herzig and Paweł Krzysztof Nowak and Thomas Müller and Francesco Piccinno and Julian Martin Eisenschlos},
75
      year={2020},
76
      eprint={2004.02349},
77
      archivePrefix={arXiv},
78
      primaryClass={cs.IR}
79
}
80
```
81
82
```bibtex
83
@misc{eisenschlos2020understanding,
84
      title={Understanding tables with intermediate pre-training}, 
85
      author={Julian Martin Eisenschlos and Syrine Krichene and Thomas Müller},
86
      year={2020},
87
      eprint={2010.00571},
88
      archivePrefix={arXiv},
89
      primaryClass={cs.CL}
90
}
91
```
92
93
```bibtex
94
@article{DBLP:journals/corr/abs-1709-00103,
95
  author    = {Victor Zhong and
96
               Caiming Xiong and
97
               Richard Socher},
98
  title     = {Seq2SQL: Generating Structured Queries from Natural Language using
99
               Reinforcement Learning},
100
  journal   = {CoRR},
101
  volume    = {abs/1709.00103},
102
  year      = {2017},
103
  url       = {http://arxiv.org/abs/1709.00103},
104
  archivePrefix = {arXiv},
105
  eprint    = {1709.00103},
106
  timestamp = {Mon, 13 Aug 2018 16:48:41 +0200},
107
  biburl    = {https://dblp.org/rec/journals/corr/abs-1709-00103.bib},
108
  bibsource = {dblp computer science bibliography, https://dblp.org}
109
}
110
```