sangmichaelxie
commited on
Commit
•
49ebabe
1
Parent(s):
b17b9fc
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,72 @@
|
|
1 |
---
|
2 |
license: mit
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
size_categories:
|
6 |
+
- 10M<n<100M
|
7 |
---
|
8 |
+
# Dataset Card for DSIR-filtered-pile-50M
|
9 |
+
|
10 |
+
## Dataset Description
|
11 |
+
|
12 |
+
- **Repository:**
|
13 |
+
- **Paper:**
|
14 |
+
- **Point of Contact: Sang Michael Xie <xie@cs.stanford.edu>**
|
15 |
+
|
16 |
+
### Dataset Summary
|
17 |
+
|
18 |
+
This dataset is a subset of The Pile, selected via the DSIR data selection method. The target distribution for DSIR is the Wikipedia and book-related subsets of The Pile.
|
19 |
+
|
20 |
+
|
21 |
+
### Languages
|
22 |
+
|
23 |
+
English (EN)
|
24 |
+
|
25 |
+
## Dataset Structure
|
26 |
+
|
27 |
+
### Data Instances
|
28 |
+
|
29 |
+
[More Information Needed]
|
30 |
+
|
31 |
+
### Data Fields
|
32 |
+
|
33 |
+
[More Information Needed]
|
34 |
+
|
35 |
+
## Dataset Creation
|
36 |
+
We first select 102.4M examples then concatenate every two examples to create 51.2M examples.
|
37 |
+
This ensures that the examples are long enough for a max token length of 512 without much padding.
|
38 |
+
We train the importance weight estimator for DSIR from The Pile validation set, where the target is Wikipedia + BookCorpus2 + Gutenberg + Books3 and the raw data come from the rest of the data sources in The Pile.
|
39 |
+
We first select 98.4M examples from non-Wikipedia and book data, then randomly select 2M from Wikipedia and 0.66M each from BookCorpus2, Gutenberg, and Books3.
|
40 |
+
After this, we concatenate every two examples.
|
41 |
+
|
42 |
+
### Source Data
|
43 |
+
The Pile
|
44 |
+
|
45 |
+
#### Initial Data Collection and Normalization
|
46 |
+
We select data from The Pile, which comes in 30 random chunks. We reserve chunk 0 for validation purposes and only consider the last 29 chunks.
|
47 |
+
We first divided the documents in The Pile into chunks of 128 words, according to whitespace tokenization.
|
48 |
+
These chunks define the examples that we do data selection on, totaling 1.7B examples.
|
49 |
+
Before DSIR, we first apply a manual quality filter (see paper for details) and only consider the examples that pass the filter.
|
50 |
+
|
51 |
+
|
52 |
+
### Annotations
|
53 |
+
|
54 |
+
There are no annotations.
|
55 |
+
|
56 |
+
|
57 |
+
## Considerations for Using the Data
|
58 |
+
|
59 |
+
The dataset is biased towards choosing data from non-Wikipedia and non-Books sources. A balanced approach would be to mix in more data from Wikipedia and books.
|
60 |
+
|
61 |
+
### Dataset Curators
|
62 |
+
|
63 |
+
Sang Michael Xie
|
64 |
+
|
65 |
+
### Citation Information
|
66 |
+
|
67 |
+
@article{xie2023data,
|
68 |
+
author = {Sang Michael Xie and Shibani Santurkar and Tengyu Ma and Percy Liang},
|
69 |
+
journal = {TODO},
|
70 |
+
title = {Data Selection for Language Models via Importance Resampling},
|
71 |
+
year = {2023},
|
72 |
+
}
|