agentlans commited on
Commit
5dfcb23
1 Parent(s): 7c30a4b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +25 -13
README.md CHANGED
@@ -1,20 +1,32 @@
1
- ## High-Quality English Sentences
2
-
3
- ### Dataset Description
 
 
 
 
 
 
 
 
 
 
4
 
5
  This dataset contains a collection of high-quality English sentences sourced from [C4](https://huggingface.co/datasets/allenai/c4) and [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (*not* [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)). The sentences have been carefully filtered and processed to ensure quality and uniqueness.
6
 
7
- ### Source Data
 
 
8
 
9
  Before filtering:
10
  - **C4:** 1 million sentences
11
  - **FineWeb:** 1 million sentences
12
 
13
- ### Data Processing
14
 
15
- 1. **Initial Filtering:**
16
  - Quality score > 0.5 with [agentlans/deberta-v3-xsmall-quality](https://huggingface.co/agentlans/deberta-v3-xsmall-quality) classifier
17
- - Sentence length >= 20 characters
18
 
19
  2. **Additional Filtering:**
20
  - Removed sentences not starting with a capital letter
@@ -22,17 +34,17 @@ Before filtering:
22
 
23
  3. **Deduplication:**
24
  - Combined sentences from both sources
25
- - Filtered for unique sentences
26
 
27
  4. **Train-Test Split:**
28
  - Train set: 90% of data
29
  - Test set: 10% of data
30
 
31
- ### Data Fields
32
 
33
  No data field. Just one sentence per line as a string.
34
 
35
- ### Data Splits
36
 
37
  | Split | Number of Sentences |
38
  |---------|---------------------|
@@ -40,13 +52,13 @@ No data field. Just one sentence per line as a string.
40
  | Test | 170 522 |
41
  | Total | 1 705 221 |
42
 
43
- ### Dataset Creation
44
 
45
  - **Curation Rationale:** To create a high-quality dataset of diverse English sentences for various NLP tasks.
46
  - **Source Contributions:** Equal initial contributions from C4 and FineWeb, with subsequent filtering and processing.
47
 
48
- ### Considerations for Using the Data
49
 
50
  - The dataset deliberately uses FineWeb and C4 instead of FineWeb-Edu in order to broaden beyond educational and academic content. However, it's still biased towards those genres.
51
  - Titles and citations are considered "sentences".
52
- - The filtering process may have introduced some biases, such as favouring more formal writing styles and omitting lists due to the capitalization requirement.
 
1
+ ---
2
+ license: odc-by
3
+ task_categories:
4
+ - text-classification
5
+ - text-generation
6
+ - feature-extraction
7
+ - sentence-similarity
8
+ language:
9
+ - en
10
+ ---
11
+ # High-Quality English Sentences
12
+
13
+ ## Dataset Description
14
 
15
  This dataset contains a collection of high-quality English sentences sourced from [C4](https://huggingface.co/datasets/allenai/c4) and [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (*not* [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)). The sentences have been carefully filtered and processed to ensure quality and uniqueness.
16
 
17
+ "High-quality" means they're legible English and not spam, although they may still have spelling and grammar errors.
18
+
19
+ ## Source Data
20
 
21
  Before filtering:
22
  - **C4:** 1 million sentences
23
  - **FineWeb:** 1 million sentences
24
 
25
+ ## Data Processing
26
 
27
+ 1. **Initial Sentence Filtering:**
28
  - Quality score > 0.5 with [agentlans/deberta-v3-xsmall-quality](https://huggingface.co/agentlans/deberta-v3-xsmall-quality) classifier
29
+ - Length >= 20 characters
30
 
31
  2. **Additional Filtering:**
32
  - Removed sentences not starting with a capital letter
 
34
 
35
  3. **Deduplication:**
36
  - Combined sentences from both sources
37
+ - Filtered for unique sentences (exact match)
38
 
39
  4. **Train-Test Split:**
40
  - Train set: 90% of data
41
  - Test set: 10% of data
42
 
43
+ ## Data Fields
44
 
45
  No data field. Just one sentence per line as a string.
46
 
47
+ ## Data Splits
48
 
49
  | Split | Number of Sentences |
50
  |---------|---------------------|
 
52
  | Test | 170 522 |
53
  | Total | 1 705 221 |
54
 
55
+ ## Dataset Creation
56
 
57
  - **Curation Rationale:** To create a high-quality dataset of diverse English sentences for various NLP tasks.
58
  - **Source Contributions:** Equal initial contributions from C4 and FineWeb, with subsequent filtering and processing.
59
 
60
+ ## Considerations for Using the Data
61
 
62
  - The dataset deliberately uses FineWeb and C4 instead of FineWeb-Edu in order to broaden beyond educational and academic content. However, it's still biased towards those genres.
63
  - Titles and citations are considered "sentences".
64
+ - The filtering process may have introduced some biases, such as favouring more formal writing styles and omitting lists due to the capitalization requirement.