File size: 7,325 Bytes
0adf04e 30150ff 89ab7b8 7c38fdd 30150ff c38963e 30150ff 12d83cf 30150ff c38963e 30150ff c38963e 30150ff c38963e 30150ff 3645b03 554dcea c38963e 554dcea 30150ff c38963e 30150ff c38963e 30150ff d58c721 946d938 8adac26 946d938 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 |
---
license: apache-2.0
---
(Works with [Mobile-Env v3.x](https://github.com/X-LANCE/Mobile-Env/tree/v3.0).)
# WikiHow Task Set
WikiHow task set is an InfoUI interaction task set based on
[Mobile-Env](https://github.com/X-LANCE/Mobile-Env) proposed in [*Mobile-Env:
An Evaluation Platform and Benchmark for Interactive Agents in LLM
Era*](https://arxiv.org/abs/2305.08144).
[WikiHow](https://www.wikihow.com/Main-Page) is a collaborative wiki site about
various real-life tips with more than 340,000 online articles. To construct the
task set, 107,448 pages are crawled, and the dumped website data occupy about
88 GiB totally.
Several task definition templates are designed according to the functions of
WikiHow app and task definitions are instantiated through the template toolkit
in Mobile-Env. 577 tasks are sampled from the extended set, which is named the
*canonical set* (`wikihow-canonical.tar.xz`). Owing to the limit of the
budgets, only 150 tasks are tested using the proposed LLM-based agent. These
150 tasks are given in `wikihow-microcanon.tar.xz`. We call it the *canonical
subset* or the *micro canonical set*.
### Website Data Replay
The replay script for [mitmproxy](https://mitmproxy.org/) is given as
`replay_url.py`. To use this replay script, the information retrieval tool
[Pyserini](https://github.com/castorini/pyserini/) is required. Four parameters
are expected to be assigned in the script:
+ The crawled data from WikiHow website (`dumps` in `wikihow.data.tar.xz`)
+ The HTML templates used to mock the search result page (`templates` in
`wikihow.data.tar.xz`)
+ The indices for the search engine based on Pyserini (`indices-t/indices` in
`wikihow.data.tar.xz`)
+ The metadata of the crawled articles (`indices-t/docs/doc_meta.csv` in
`wikihow.data.tar.xz`)
All the required data are offered in `wikihow.data.tar.xz`. (The archive is
about 78 GiB. And the decompressed data are about 88 GiB.) The archive is split
into two pieces (`wikihow.data.tar.xz.00` and `wikihow.data.tar.xz.01`). You
can use `cat` to concatenate them:
```sh
cat wikihow.data.tar.xz.00 wikihow.data.tar.xz.01 >wikihow.data.tar.xz
```
The SHA256 checksums are provided in `wikihow.data.tar.xz.sha256` to check the
integrity.
To run the script:
```sh
mitmproxy --showhost -s replay_url.py
```
### Certificate Unpinning Plan
The `syscert` plan proposed by Mobile-Env works for WikiHow app. You can
complete the config according to the [guideline of
Mobile-Env](https://github.com/X-LANCE/Mobile-Env/blob/master/docs/dynamic-app-en.md).
The available APK package from [APKCombo](https://apkcombo.com/) is provided.
And note to use the AVD image of version Android 11.0 (API Level 30) (Google
APIs) to obtain the best compatibility and the root-enabled ADBD.
### Human-Rewritten Instructions
Human-rewritten instructions for the *canonical set* are release under
`instruction_rewriting/`. An AndroidEnv wrapper `InstructionRewritingWrapper`
is provided to load the rewritten instructions (`merged_doccano.json`) and
public patterns (`pattern-*.txt`). The annotations are collected via
[doccano](https://doccano.github.io/doccano/). The patterns are parsed by
[`sentence_pattern.py`](instruction_rewriting/sentence_pattern.py).
### Details of Sub-Tasks
WikiHow taks are crafted from 16 types of sub-tasks:
* `home2search`, instructing to search for an article from the home page.
* `search2article`, `author2article`, & `category2article`, instructing to
access an article from search result page, author information page, and
category content page, respectively.
* `article2about`, instructing to access the about page from article page.
* `article2author`, instructing to access author information page from article
page.
* `article2category`, instructing to access category content page from article
page.
* `article2reference`, instructing to check reference list on article page.
* `article2rate_no`, instructing to rate no for article
* `article2rate_yes`, instructing to rate yes for article
* `article2share`, instructing to share article
* `article2bookmark`, instructing to bookmark article and then check the
bookmarks.
* `article2steps`, crafted from `stepped_summary` questions in
[wikihow-lists](https://huggingface.co/datasets/b-mc2/wikihow_lists)
* `article2ingredientes`, crafted from `ingredients` questions in
[wikihow-lists](https://huggingface.co/datasets/b-mc2/wikihow_lists)
* `article2needed_items`, crafted from `needed_items` questions in
[wikihow-lists](https://huggingface.co/datasets/b-mc2/wikihow_lists)
* `article2summary`, crafted from
[WikiHowNFQA](https://huggingface.co/datasets/Lurunchik/WikiHowNFQA) tasks
A template is composed for each sub-task, containing a group of filling slots
expecting some keywords like article title, author name, question, and
groundtruth answer. Then these keywords are sampled from the crawled app data
or from the two QA datasets to instantiate the templates. Subsequently, the
instantiated templates are concatenated into multi-stage task definitions under
the constraint that the target page/element/answer (the part after `2`, *e.g.*,
`share` from `article2share`) is directly on/referenced by the current page
(the part before `2`, *e.g.*, `article` from `article2share`). Finally, we
obtained the task set of 150 multistage tasks in which there are 2.68
single-stage sub-tasks averagely.
The multistage tasks containing different sub-tasks are suffixed with different
numbers. The meanings of suffixes and the number of suffixed tasks in the micro
canonical set are list in the following table:
| Suffix | Sub-tasks | #Tasks |
|--------|------------------------------------------|--------|
| 0 | `home-search-article-about` | 18 |
| 1 | `home-search-article-rate_no` | 6 |
| 2 | `home-search-article-rate_yes` | 10 |
| 3 | `home-search-article-share` | 11 |
| 4 | `home-search-article-author[-article]` | 7 |
| 5 | `home-search-article-bookmark` | 13 |
| 6 | `home-search-article-category[-article]` | 9 |
| 7 | `home-search-article-reference` | 11 |
| 8 | `home-search-article` | 25 |
| 9 | `home-search-steps` | 15 |
| 10 | `home-search-needed_items` | 10 |
| 11 | `home-search-ingredients` | 5 |
| 12 | `home-search-summary` | 10 |
### About
This task set is developed and maintained by [SJTU
X-Lance](https://x-lance.sjtu.edu.cn/en). The corresponding paper is available
at <https://arxiv.org/abs/2305.08144>.
If you find WikiHow task set useful in your research, you can cite the project
using the following BibTeX:
```bibtex
@article{DanyangZhang2023_MobileEnv_WikiHow,
title = {{Mobile-Env}: An Evaluation Platform and Benchmark for LLM-GUI Interaction},
author = {Danyang Zhang and
Lu Chen and
Zihan Zhao and
Ruisheng Cao and
Kai Yu},
journal = {CoRR},
volume = {abs/2305.08144},
year = {2023},
url = {https://arxiv.org/abs/2305.08144},
eprinttype = {arXiv},
eprint = {2305.08144},
}
```
|