---
license: odc-by
tags:
  - gsa
  - gist-sparse-attention
  - long-context
  - pretraining
source_datasets:
  - HuggingFaceFW/fineweb
---

# GSA-PT-Llama-3.2-1B-chunk8-data

This is the **continue pretraining dataset** used for training GSA (Gist Sparse Attention) models with chunk size `chunk8`.

## Dataset Description

- **Source**: Filtered from [FineWeb (sample-10BT)](https://huggingface.co/datasets/HuggingFaceFW/fineweb/viewer/sample-10BT)
- **Filtering**: Samples with sequence lengths ranging from **2K to 4K** tokens
- **Chunk size**: `chunk8`
- **Splits**: `train`, `eval`

Each sample is tokenized and formatted with GSA gist tokens for continued pretraining.

## Paper

**GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding**

## Related Models

- [yuzhenm/GSA-PT-Llama-3.2-1B-chunk8](https://huggingface.co/yuzhenm/GSA-PT-Llama-3.2-1B-chunk8) — model trained on this dataset