Update README.md

README.md

@@ -21,3 +21,53 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+
+# Description
+
+This dataset is a sampled subset of the [Pile](https://huggingface.co/datasets/EleutherAI/pile) dataset. 
+We used [DSIR](https://github.com/p-lambda/dsir) a data selection tool with importance resampling for subsampling.
+
+The subset sample distribution is:
+
+```json
+{
+   'Pile-CC': 198245,
+   'OpenWebText2': 122382,
+   'FreeLaw': 37517,
+   'USPTO Backgrounds': 10195,
+   'Wikipedia (en)': 8072,
+   'PubMed Central': 5849,
+   'PubMed Abstracts': 4965,
+   'Gutenberg (PG-19)': 2712,
+   'BookCorpus2': 2550,
+   'Books3': 2432,
+   'StackExchange': 1753,
+   'PhilPapers': 1560,
+   'YoutubeSubtitles': 1187,
+   'OpenSubtitles': 1015,
+   'ArXiv': 610,
+   'NIH ExPorter': 476,
+   'Enron Emails': 439,
+   'EuroParl': 419,
+   'Github': 390,
+   'HackerNews': 259
+}
+```
+
+
+The dataset contains ~100M words of text. This can be checked with:
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("PatrickHaller/dsir-pile-100M-words")
+
+count = 0
+for row in ds["train"]:
+    count += len(row["text"].split(" "))
+
+print(count)
+
+# Out: 99999861
+```