File size: 6,404 Bytes
8084f28
 
 
 
 
2e1428b
 
 
d2f54c1
 
 
 
 
 
 
 
 
 
 
 
2e1428b
d2f54c1
2e1428b
d2f54c1
2e1428b
 
d2f54c1
2e1428b
d2f54c1
2e1428b
8084f28
6a7bbfd
8084f28
 
 
 
d31acc0
6a7bbfd
 
8084f28
 
 
 
6a7bbfd
8084f28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c62aa0
8084f28
6a7bbfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2483bb0
6a7bbfd
 
 
 
 
 
 
 
 
 
 
 
2483bb0
6a7bbfd
 
 
 
 
 
 
 
 
 
 
 
 
 
8084f28
 
 
 
 
 
 
 
 
 
 
6a7bbfd
8084f28
6a7bbfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8084f28
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
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
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
---
language: en
license: mit
tags:
- exbert
datasets:
- squad_v2
thumbnail: https://thumb.tildacdn.com/tild3433-3637-4830-a533-353833613061/-/resize/720x/-/format/webp/germanquad.jpg
model-index:
- name: deepset/tinybert-6l-768d-squad2
  results:
  - task:
      type: question-answering
      name: Question Answering
    dataset:
      name: squad_v2
      type: squad_v2
      config: squad_v2
      split: validation
    metrics:
    - type: exact_match
      value: 73.8248
      name: Exact Match
      verified: true
      verifyToken: eyJhbGciOiJFZERTQSIsInR5cCI6IkpXVCJ9.eyJoYXNoIjoiMGFmZmFiN2E5ODZkOTkyMjQ1NTUzMmQwMjc0M2RlYzVlNmM4YTFlNzA4YzIwY2JkY2EyNDg2ZTY3OTdjZTVlZiIsInZlcnNpb24iOjF9.ZZ6c2OI3lzeNhuSWTh28j00zk-sPrqkTvdVBZv2wJc1D4YnR-xOj72haybT6MV_xeYqTg3-x9L8PsWSS20NaDw
    - type: f1
      value: 77.1684
      name: F1
      verified: true
      verifyToken: eyJhbGciOiJFZERTQSIsInR5cCI6IkpXVCJ9.eyJoYXNoIjoiNzAxMDk1YzI5ZjA2N2ZmMzAxNjgxYzJiNzAzYmI1ZWU5ZDRmYWY3OWJmMjlmNDcyMGE0YWY5NjNhZTk4YWY5ZSIsInZlcnNpb24iOjF9.rF3raNGUSYv5D2xzWLZztD99vwDKvWb22LG32RomrDGP6XKTbCVqZzAw5UFw93jKb0VoLApbQQ-AOGxLj3U_Cg
---
# tinybert for Extractive QA 

## Overview
**Language model:** deepset/tinybert-6L-768D-squad2   
**Language:** English  
**Training data:** SQuAD 2.0 training set x 20 augmented + SQuAD 2.0 training set without augmentation  
**Eval data:** SQuAD 2.0 dev set 
**Code:**  See [an example extractive QA pipeline built with Haystack](https://haystack.deepset.ai/tutorials/34_extractive_qa_pipeline)  
**Infrastructure**: 1x V100 GPU  
**Published**: Dec 8th, 2021

## Details
- Haystack's intermediate layer and prediction layer distillation features were used for training (based on [TinyBERT](https://arxiv.org/pdf/1909.10351.pdf)). deepset/bert-base-uncased-squad2 was used as the teacher model and huawei-noah/TinyBERT_General_6L_768D was used as the student model.

## Hyperparameters
### Intermediate layer distillation
```
batch_size = 26
n_epochs = 5
max_seq_len = 384
learning_rate = 5e-5
lr_schedule = LinearWarmup
embeds_dropout_prob = 0.1
temperature = 1
```
### Prediction layer distillation 
```
batch_size = 26
n_epochs = 5
max_seq_len = 384
learning_rate = 3e-5
lr_schedule = LinearWarmup
embeds_dropout_prob = 0.1
temperature = 1
distillation_loss_weight = 1.0
```

## Usage

### In Haystack
Haystack is an AI orchestration framework to build customizable, production-ready LLM applications. You can use this model in Haystack to do extractive question answering on documents. 
To load and run the model with [Haystack](https://github.com/deepset-ai/haystack/):
```python
# After running pip install haystack-ai "transformers[torch,sentencepiece]"

from haystack import Document
from haystack.components.readers import ExtractiveReader

docs = [
    Document(content="Python is a popular programming language"),
    Document(content="python ist eine beliebte Programmiersprache"),
]

reader = ExtractiveReader(model="deepset/tinybert-6l-768d-squad2")
reader.warm_up()

question = "What is a popular programming language?"
result = reader.run(query=question, documents=docs)
# {'answers': [ExtractedAnswer(query='What is a popular programming language?', score=0.5740374326705933, data='python', document=Document(id=..., content: '...'), context=None, document_offset=ExtractedAnswer.Span(start=0, end=6),...)]}
```
For a complete example with an extractive question answering pipeline that scales over many documents, check out the [corresponding Haystack tutorial](https://haystack.deepset.ai/tutorials/34_extractive_qa_pipeline).

### In Transformers
```python
from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline

model_name = "deepset/tinybert-6l-768d-squad2"

# a) Get predictions
nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)
QA_input = {
    'question': 'Why is model conversion important?',
    'context': 'The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks.'
}
res = nlp(QA_input)

# b) Load model & tokenizer
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
```

## Performance
```
"exact": 71.87736882001179
"f1": 76.36111895973675
```

## Authors
- Timo M枚ller: `timo.moeller [at] deepset.ai`
- Julian Risch: `julian.risch [at] deepset.ai`
- Malte Pietsch: `malte.pietsch [at] deepset.ai`
- Michel Bartels: `michel.bartels [at] deepset.ai`

## About us

<div class="grid lg:grid-cols-2 gap-x-4 gap-y-3">
    <div class="w-full h-40 object-cover mb-2 rounded-lg flex items-center justify-center">
         <img alt="" src="https://raw.githubusercontent.com/deepset-ai/.github/main/deepset-logo-colored.png" class="w-40"/>
     </div>
     <div class="w-full h-40 object-cover mb-2 rounded-lg flex items-center justify-center">
         <img alt="" src="https://raw.githubusercontent.com/deepset-ai/.github/main/haystack-logo-colored.png" class="w-40"/>
     </div>
</div>

[deepset](http://deepset.ai/) is the company behind the production-ready open-source AI framework [Haystack](https://haystack.deepset.ai/).

Some of our other work: 
- [Distilled roberta-base-squad2 (aka "tinyroberta-squad2")](https://huggingface.co/deepset/tinyroberta-squad2)
- [German BERT](https://deepset.ai/german-bert), [GermanQuAD and GermanDPR](https://deepset.ai/germanquad), [German embedding model](https://huggingface.co/mixedbread-ai/deepset-mxbai-embed-de-large-v1)
- [deepset Cloud](https://www.deepset.ai/deepset-cloud-product), [deepset Studio](https://www.deepset.ai/deepset-studio)

## Get in touch and join the Haystack community

<p>For more info on Haystack, visit our <strong><a href="https://github.com/deepset-ai/haystack">GitHub</a></strong> repo and <strong><a href="https://docs.haystack.deepset.ai">Documentation</a></strong>. 

We also have a <strong><a class="h-7" href="https://haystack.deepset.ai/community">Discord community open to everyone!</a></strong></p>

[Twitter](https://twitter.com/Haystack_AI) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Discord](https://haystack.deepset.ai/community) | [GitHub Discussions](https://github.com/deepset-ai/haystack/discussions) | [Website](https://haystack.deepset.ai/) | [YouTube](https://www.youtube.com/@deepset_ai)

By the way: [we're hiring!](http://www.deepset.ai/jobs)