Update README.md

README.md

@@ -15,3 +15,18 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# OpenWebTextCorpus tokenized for Gemma
+
+This dataset is a pre-tokenized version of the [Skylion007/openwebtext](https://huggingface.co/datasets/Skylion007/openwebtext) dataset
+using the [gemma](https://huggingface.co/google/gemma-2b) tokenizer. As such, this dataset follows the same licensing as the original openwebtext dataset.
+
+This pre-tokenization is done as a performance optimization for using the openwebtext dataset with a Gemma model (gemma-2b, gemma-2b-it, gemma-7b, gemma-7b-it).
+This dataset was created using [SAELens](https://github.com/jbloomAus/SAELens), with the following settings:
+
+- context_size: 8192
+- shuffled: true
+- begin_batch_token: "bos",
+- begin_sequence_token: null,
+- sequence_separator_token: "bos"
+- sae_lens_version: "3.3.0"