sehunnnn commited on
Commit
be817d0
1 Parent(s): 8bee08a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +97 -10
README.md CHANGED
@@ -1,20 +1,107 @@
1
  ---
2
- license: cc-by-4.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  task_categories:
 
4
  - text-generation
5
  - question-answering
 
 
6
  - table-question-answering
7
  language:
8
  - en
9
  tags:
10
- - SQL
11
- - code
12
- - NLP
13
- - text-to-sql
14
- - context-sql
15
- - spider
16
- - wikisql
17
- - sqlglot
18
- pretty_name: sql-create-context
19
  size_categories:
20
  - 10K<n<100K
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ dataset_info:
3
+ features:
4
+ - name: input
5
+ dtype: string
6
+ - name: output
7
+ dtype: string
8
+ - name: source
9
+ dtype: string
10
+ - name: subset
11
+ dtype: string
12
+ splits:
13
+ - name: train
14
+ num_bytes: 63759065
15
+ num_examples: 23652
16
+ - name: validation
17
+ num_bytes: 6190242
18
+ num_examples: 2042
19
+ - name: test
20
+ num_bytes: 6080212
21
+ num_examples: 2045
22
+ download_size: 45525146
23
+ dataset_size: 76029519
24
  task_categories:
25
+ - text2text-generation
26
  - text-generation
27
  - question-answering
28
+ - conversational
29
+ - summarization
30
  - table-question-answering
31
  language:
32
  - en
33
  tags:
34
+ - instruction-tuning
35
+ pretty_name: longform
 
 
 
 
 
 
 
36
  size_categories:
37
  - 10K<n<100K
38
+ ---
39
+ # LongForm
40
+ The LongForm dataset is created by leveraging English corpus
41
+ examples with augmented instructions. We select a
42
+ diverse set of human-written
43
+ documents from existing corpora such as C4 and
44
+ Wikipedia and generate instructions for the given
45
+ documents via LLMs. Then, we extend these examples with structured corpora examples such as Stack Exchange and WikiHow and task examples such as question answering, email writing, grammar error correction, story/poem generation, and text summarization.
46
+
47
+ ## Distribution
48
+ The distribution of the LongForm dataset in terms of the source of examples is below. It contains examples generated from raw text corpora via LLMs, structured corpus examples, as well as various NLP task examples such as email writing, grammar error correction, story/poem generation, and text summarization.
49
+ | **Type** | **Source** | **Number of Examples** |
50
+ |------------------------|----------------|------------------------|
51
+ | **Corpora** | C4 | 10,000 |
52
+ | | Wikipedia | 5,000 |
53
+ | **Structured Corpora** | Stack Exchange | 4,380 |
54
+ | | WikiHow | 2,500 |
55
+ | **Tasks** | NIv2 | 3,684 |
56
+ | | Big Bench | 600 |
57
+ | | BEA-GEC | 1,203 |
58
+ | | Enron | 372 |
59
+ | **Total** | | 27,739 |
60
+ | | | |
61
+ | **Train** | | 23,652 |
62
+ | **Validation** | | 2,042 |
63
+ | **Test** | | 2,045 |
64
+
65
+ ## Models
66
+ | | **All** | **Recipe Generation** | **ELI5** | **Writing Prompts** |
67
+ |-----------------------|---------|-----------------------------------|----------|---------------------|
68
+ | **T0++** | 10.9 | 18.7 | 3.8 | 10.2 |
69
+ | **Tk-Instruct** | 6.3 | 12.9* | 3.6 | 2.4 |
70
+ | **Flan-T5** | 10.6 | 20.9* | 3.5 | 7.4 |
71
+ | **Alpaca-LLaMA-7B** | 14.6 | 19.5 | 12.5 | 11.8 |
72
+ | **OPT-30B** | 11.1 | 18.6 | 12.2 | 2.6 |
73
+ | **[LongForm-T5-XL](https://huggingface.co/akoksal/LongForm-T5-XL)** | 16.3 | 20.2 | 18.3 | 10.6 |
74
+ | **[LongForm-OPT-2.7B](https://huggingface.co/akoksal/LongForm-OPT-2.7B)** | 17.8 | 15.5 | 17.9 | **19.9** |
75
+ | **[LongForm-OPT-6.7B](https://huggingface.co/akoksal/LongForm-OPT-6.7B)** | 17.7 | 16.9 | 17.2 | 19.0 |
76
+ | [**LongForm-LLaMA-7B**](https://huggingface.co/akoksal/LongForm-LLaMA-7B-diff)‡ | **19.7** | **21.7** | **18.6** | 18.9 |
77
+
78
+ ‡: We can just release the difference between LongForm-LLaMA-7B and pretrained LLaMA-7B publicly due to restrictions of LLaMA models.
79
+
80
+ ## Examples
81
+ | **Source** | **Instruction** | **Output**|
82
+ |------------------------|----------------|------------------------|
83
+ | C4 | What is the likelihood that Donald Trump's HHS secretary nominee, Alex Azar, will push policies that benefit pharmaceutical companies' pocketbooks? | His record suggests he’s unlikely to push pricing policies that would hit pharmaceutical companies’ pocketbooks.<br />President Donald Trump tweeted Monday that his nominee for HHS secretary, Alex Azar will “be a star for … lower drug prices!” But the record of the former top executive for Eli Lilly, which tripled the price of a top-selling insulin drug while he led its U.S. operation, suggests a different story.<br />Lilly is one of three drug companies targeted by a class-action lawsuit that accuses the company[...]<br />“This nominee has impeccable big pharma credentials but non-existent credentials when it comes to concerns on drug prices,” said Peter Welch, co-chair of the House Democrats Drug Pricing Task Force.
84
+ |Wikipedia | "Where is San Saba located in Rome, Italy?" Respond in 2 sentences. |San Saba is an ancient basilica church in Rome, Italy. It lies on the so-called "Piccolo Aventino", which is an area close to the ancient Aurelian Walls next to the Aventine Hill and Caelian Hill.|
85
+
86
+ ## Dataset Description
87
+
88
+ - **Repository:** https://github.com/akoksal/LongForm
89
+ - **Paper:** https://arxiv.org/abs/2304.08460
90
+ - **Version:** v1.0 - April 18, 2023
91
+ - **Contact:** [Abdullatif Köksal](https://twitter.com/akoksal_)
92
+
93
+ ## License
94
+ The LongForm project is subject to a MIT License with custom limitations for restrictions imposed by OpenAI (for the instruction generation part), as well as the license of language models (OPT, LLaMA, and T5).
95
+
96
+
97
+ ## Citation
98
+ ```
99
+ @misc{koksal2023longform,
100
+ title={LongForm: Optimizing Instruction Tuning for Long Text Generation with Corpus Extraction},
101
+ author={Abdullatif Köksal and Timo Schick and Anna Korhonen and Hinrich Schütze},
102
+ year={2023},
103
+ eprint={2304.08460},
104
+ archivePrefix={arXiv},
105
+ primaryClass={cs.CL}
106
+ }
107
+ ```