Updating README

README.md

@@ -24,4 +24,25 @@ task_categories:
 language:
 - en
 pretty_name: Pretokenized Paloma
----
+---
+
+# The Pretokenized Paloma Benchmark Dataset
+
+This dataset is a compact, pre-tokenized evaluation dataset designed to complement the [pretokenized-dolma](https://huggingface.co/datasets/pico-lm/pretokenized-dolma) training set. Built from the [Paloma corpus](https://github.com/allenai/OLMo-Eval/blob/main/paloma/README.md) (Allen Institute), this benchmark was designed to not contain any data overlap with Dolma and is ideal for evaluating models trained on it.
+
+### Overview 
+
+Features:
+- Pre-tokenized with the same tokenizer as pretokenized-dolma: [allenai/OLMo-7B-0724-hf](https://huggingface.co/allenai/OLMo-7B-0724-hf)
+- Sequence length: 2048 tokens
+- Ideal for perplexity calculations for models trained on pretokenized-dolma
+
+We release the exact scripts we use to create this dataset in our [pico-lm/pico-dataset](https://github.com/pico-lm/pico-dataset) GitHub repo.
+
+### Usage
+
+```
+from datasets import load_dataset
+dataset = load_dataset("pico-lm/pretokenized-paloma", streaming=True)
+```
+