agentlans commited on
Commit
51afb2f
·
verified ·
1 Parent(s): e5a7195

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +41 -1
README.md CHANGED
@@ -68,4 +68,44 @@ language:
68
  - en
69
  tags:
70
  - pretraining
71
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
  - en
69
  tags:
70
  - pretraining
71
+ ---
72
+ # High Quality Text Dataset
73
+
74
+ A curated collection of English-language texts for AI training and research.
75
+
76
+ ### Sources
77
+ - [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
78
+ - [openbmb/Ultra-FineWeb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb)
79
+ - [Zyphra/Zyda-2](https://huggingface.co/datasets/Zyphra/Zyda-2)
80
+
81
+ Each dataset was
82
+ 1. Split into ~2 000-token chunks using the [LLaMA 3.1 tokenizer](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct)
83
+ 2. Cleaned (normalized spaces, punctuation, and characters, replaced emails and phone numbers with placeholders)
84
+ 3. Scored using the [`agentlans/GIST-all-MiniLM-L6-v2-quality-v3`](https://huggingface.co/agentlans/GIST-all-MiniLM-L6-v2-quality-v3) classifier
85
+ 4. Exact duplicates removed
86
+
87
+ 100 000 chunks per source were included.
88
+
89
+ ### Clustering
90
+ Agglomerative clustering was applied using embeddings from the [`Snowflake/snowflake-arctic-embed-xs`](https://huggingface.co/Snowflake/snowflake-arctic-embed-xs)
91
+ model at multiple cluster counts: 100, 200, 500, 1 000, 2 000, 5 000, 10 000, 20 000, 50 000, 100 000, and 200 000 clusters, enabling flexible dataset configurations.
92
+
93
+ ### Example Entry
94
+
95
+ ```json
96
+ {
97
+ "text": "Dr. Louise Glew has been appointed the Global Lead Scientist for WWF's Global Science Team. Louise's research focuses on understanding the social and ecological impacts of conservation interventions [...]",
98
+ "quality": 2.0699,
99
+ "source": "openbmb/Ultra-FineWeb"
100
+ }
101
+ ```
102
+
103
+ ### Limitations
104
+ - Focused on academic, educational, and pedagogical content for a general audience.
105
+ - May contain outdated or unreliable information (including self-published material, pseudoscience, conspiracy theories).
106
+ - Quality scores reflect syntax and tone, not factual accuracy.
107
+ - Some repetition may occur (for example, dictionary entries, geographic distance calculations).
108
+ - The entries may be broken in the middle of a word or sentence.
109
+
110
+ ### License
111
+ Provided under the Open Data Commons Attribution License (ODC-BY).