Hugging Face's logo

Datasets:
openbmb
/
Ultra-FineWeb

Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Chinese
Size:
1B - 10B
ArXiv:
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files Community
17
BigDong commited on
Commit
04fabdc
·
1 Parent(s): 1b0db84

update readme

Browse files
Files changed (1) hide show
  1. README.md +7 -7
README.md CHANGED
@@ -70,18 +70,18 @@ We utilize the MiniCPM-1.2B model architecture with the MiniCPM3-4B tokenizer. E
70
  Detailed evaluation results are reported below:
71
 
72
  - **Individual data experiments.** We perform isolated training runs using single datasets, facilitating direct comparisons between differently processed data from identical sources.
73
- <img src="assets/individual-english-table.png" alt="Performance Scatter Plot" width="75%">
74
- <img src="assets/individual-chinese-table.png" alt="Performance Scatter Plot" width="75%">
75
- <img src="assets/individual-plot.png" alt="Performance Scatter Plot" width="100%">
76
 
77
  - **Mixed Data Experiments.** We use a mix of 60% English data, 30% Chinese data, and 10% code data (StarCoder-v2).
78
- <img src="assets/mix-table.png" alt="Performance Scatter Plot" width="75%">
79
- <img src="assets/mix-plot.png" alt="Performance Scatter Plot" width="100%">
80
 
81
  - **Loss and Performance Estimation Results.** We use the performance estimation methods proposed in [Densing Law](https://arxiv.org/abs/2412.04315) for further analysis and verification of the effectiveness of Ultra-FineWeb.
82
 
83
- <img src="assets/densing-law-table.png" alt="Performance Scatter Plot" width="75%">
84
- <img src="assets/densing-law-plot.png" alt="Performance Scatter Plot" width="100%">
85
 
86
  ## ❤️ Acknowledgements
87
 
 
70
  Detailed evaluation results are reported below:
71
 
72
  - **Individual data experiments.** We perform isolated training runs using single datasets, facilitating direct comparisons between differently processed data from identical sources.
73
+ <img src="assets/individual-english-table.png" alt="Individual English Table" width="75%">
74
+ <img src="assets/individual-chinese-table.png" alt="Individual Chinese Table" width="75%">
75
+ <img src="assets/individual-plot.png" alt="Individual Plot" width="100%">
76
 
77
  - **Mixed Data Experiments.** We use a mix of 60% English data, 30% Chinese data, and 10% code data (StarCoder-v2).
78
+ <img src="assets/mix-table.png" alt="Mix Table" width="75%">
79
+ <img src="assets/mix-plot.png" alt="Mix Plot" width="100%">
80
 
81
  - **Loss and Performance Estimation Results.** We use the performance estimation methods proposed in [Densing Law](https://arxiv.org/abs/2412.04315) for further analysis and verification of the effectiveness of Ultra-FineWeb.
82
 
83
+ <img src="assets/densing-law-table.png" alt="Densing Law Table" width="75%">
84
+ <img src="assets/densing-law-plot.png" alt="Densing Law Plot" width="100%">
85
 
86
  ## ❤️ Acknowledgements
87