---
license: apache-2.0
task_categories:
- text-generation
---

# Prolong_64K_v2_Llama2_Tokenizer

This is the Prolong_64K dataset, tokenized using the [Llama-2-7b-hf tokenizer](https://github.com/microsoft/Samba/blob/main/scripts/prepare_slimpajama.py#L22) for use in Samba-style training.

This dataset was used in the research paper: [Rethinking Language Model Scaling under Transferable Hypersphere Optimization](https://huggingface.co/papers/2603.28743).

The official training codebase can be found at [GitHub - microsoft/ArchScale](https://github.com/microsoft/ArchScale).

## Download

👉 You can download and unzip the dataset from: [prolong_64K_v2.zip](https://huggingface.co/datasets/jsun/Prolong_64K_v2_Llama2_Tokenizer/blob/main/prolong_64K_v2.zip)

```bash
wget -c https://huggingface.co/datasets/jsun/Prolong_64K_v2_Llama2_Tokenizer/resolve/main/prolong_64K_v2.zip -O prolong_64K_v2.zip
sudo apt install zip # Ubuntu
unzip prolong_64K_v2.zip -d prolong_64K_v2
```

## Usage

Once extracted, the dataset can be loaded using the [PackedDataset](https://github.com/microsoft/Samba/blob/383c016f2fb20ce75eed777761e1a4456c87b2b0/lit_gpt/packed_dataset.py#L33) class from the Samba/ArchScale codebase.

Example training scripts utilizing this data format are provided in the [ArchScale repository](https://github.com/microsoft/ArchScale).