permutans commited on
Commit
06a77a8
·
verified ·
1 Parent(s): d54d86f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +82 -1
README.md CHANGED
@@ -58,4 +58,85 @@ tags:
58
  pretty_name: FineWeb BBC News
59
  size_categories:
60
  - 1M<n<10M
61
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  pretty_name: FineWeb BBC News
59
  size_categories:
60
  - 1M<n<10M
61
+ ---
62
+
63
+ # Dataset Card for BBC News from FineWeb
64
+
65
+ This dataset provides a filtered subset of BBC News articles from each subset of the FineWeb dataset, expected to contain approximately 300M articles from BBC News domains.
66
+
67
+ ## Dataset Details
68
+
69
+ ### Dataset Description
70
+
71
+ - **Curated by:** Louis Maddox (@permutans on HuggingFace and X/Twitter)
72
+ - **License:** ODC-BY (inherited from FineWeb)
73
+ - **Language:** English
74
+
75
+ ### Dataset Sources
76
+ - **Repository:** https://huggingface.co/datasets/permutans/fineweb-bbc-news
77
+ - **Source Dataset:** HuggingFaceFW/fineweb
78
+ - **Paper:** https://arxiv.org/abs/2406.17557 (FineWeb paper)
79
+
80
+ ## Uses
81
+
82
+ ### Direct Use
83
+ Suitable for text analysis and NLP tasks focused on news content, particularly when working with BBC News articles. The dataset provides cleaned article text without metadata like bylines or publication dates.
84
+
85
+ ### Out-of-Scope Use
86
+ This dataset should not be used as a comprehensive archive of BBC News content, as it represents only articles captured in FineWeb's crawls. It should not be assumed to contain all articles from any given time period.
87
+
88
+ ## Dataset Structure
89
+
90
+ ### Data Instances
91
+ Example format:
92
+ ```python
93
+ {
94
+ 'url': 'news.bbc.co.uk/news/article-path',
95
+ 'text': 'Article content...'
96
+ }
97
+ ```
98
+
99
+ ### Data Fields
100
+ - `url`: URL of the article with query parameters removed
101
+ - `text`: Full article text content
102
+
103
+ ### Data Statistics
104
+ - Contains approximately 300k articles per subset
105
+
106
+ ## Dataset Creation
107
+
108
+ ### Curation Rationale
109
+ Created to provide an easily accessible dataset of BBC news articles while offering a focused view into the FineWeb dataset's coverage of major news sources. Enables analysis of FineWeb's completeness and motivates investigation of alternative data acquisition methods.
110
+
111
+ ### Source Data
112
+ #### Data Collection and Processing
113
+ - Filtered from FineWeb's subsets
114
+ - Limited to domains: news.bbc.co.uk, www.bbc.co.uk/news, www.bbc.com/news
115
+ - URL cleaning: removed query parameters
116
+ - Regional news content may be excluded (under regional subsite paths rather than `/news`)
117
+ - No modifications to article text content
118
+
119
+ #### Personal and Sensitive Information
120
+ Article texts contain only the main content body, without bylines or metadata.
121
+
122
+ ## Bias, Risks, and Limitations
123
+
124
+ - No validation split in current version
125
+ - Original publication dates not available (FineWeb timestamps were crawl dates)
126
+ - Section/index pages not yet filtered out from article pages
127
+ - Regional news content explicitly excluded due to sparse coverage
128
+ - Relationship between news.bbc.co.uk and bbc.co.uk/news domains needs investigation
129
+ - Coverage may be incomplete compared to full BBC News archive
130
+
131
+ ### Recommendations
132
+ Users should be aware that this represents a subset of BBC News content. For applications requiring comprehensive coverage or accurate publication dates, additional data sources should be considered.
133
+
134
+ ## Future Directions
135
+ - Addition of publication dates through targeted crawling
136
+ - Filtering to distinguish between section pages and article pages
137
+
138
+ ## Citation
139
+ Please cite the original FineWeb dataset when using this data. A reference to this one would be welcome but not necessary, I consider this a derivative work.
140
+
141
+ ## Dataset Card Authors
142
+ Louis Maddox (@permutans)