---
license: apache-2.0
language:
- en
pretty_name: ElephantBench Source Corpus
size_categories:
- 10M<n<100M
---

# ElephantBench Source Corpus

This dataset contains the low-quality partition ($D_{\mathrm{low}}$) used to construct [ElephantBench](https://github.com/tencent/ElephantBench). The released benchmark is available separately at [panzs19/ElephantBench](https://huggingface.co/datasets/tencent/ElephantBench).

$D_{\mathrm{low}}$ is derived from [cx-cmu/repro-organic-data-72B](https://huggingface.co/datasets/cx-cmu/repro-organic-data-72B) using the RePro fastText quality score. It contains 47,850,862 English web documents in 600 JSONL.zstd shards (approximately 65 GiB compressed). Each record retains the source text, URL, quality score, and original metadata required by the benchmark-construction pipeline.

The files in this repository are construction inputs, not ElephantBench evaluation questions. See the [ElephantBench repository](https://github.com/tencent/ElephantBench#benchmark-construction) for the complete construction workflow.