Update README.md

README.md

@@ -15,3 +15,22 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# OpenWebTextCorpus tokenized for Llama 3
+
+This dataset is a pre-tokenized version of the [Skylion007/openwebtext](https://huggingface.co/datasets/Skylion007/openwebtext) dataset
+using the [llama3](https://huggingface.co/meta-llama/Meta-Llama-3-8B) 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 Llama3 model.
+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: "eos"
+- sae_lens_version: "3.3.0"
+
+The `eos` token was used as a separator between sequences, since this resulted in the lowest loss experimentally.
+Ideally we would like to use the same tokenization settings as used by the original Llama3 training regime, so if
+you have information that the original Llama3 was trained using a different tokenization setup, please reach out!