---
license: other
language:
  - en
task_categories:
  - question-answering
  - text-generation
size_categories:
  - 100K<n<1M
---

# ZsRE

Canonical ZsRE model-editing splits, repackaged as JSON Lines for direct use
with `datasets.load_dataset("CrowdMind/ZsRE")`. The train and test files
preserve the source records and fields without synthetic additions.

## Splits

- `train.jsonl`: 163,196 records from the ROME/MEND training split.
- `test.jsonl`: 19,086 records from the ROME/MEND evaluation split, exposed as
  `test` here.

Each row contains `subject`, `src`, `pred`, `rephrase`, `alt`, `answers`,
`loc`, `loc_ans`, and `cond`. In this model-editing format, `src` is the edit
prompt, `alt` is the requested new answer, `answers` contains the original
answer(s), `rephrase` tests generalization, and `loc` / `loc_ans` test locality.

## Provenance and license

The split files are from the [ROME project data archive](https://rome.baulab.info/data/dsets/)
(`zsre_mend_train.json` and `zsre_mend_eval.json`). The underlying benchmark
is ZsRE from Levy et al., [“Zero-Shot Relation Extraction via Reading
Comprehension”](https://aclanthology.org/K17-1034/). The source archive makes
the data publicly downloadable but does not state a separate data license;
this repository therefore does not claim a new license for the data. Please
refer to the original sources for terms and attribution.

## Citation

```bibtex
@inproceedings{levy-etal-2017-zero,
  title = {Zero-Shot Relation Extraction via Reading Comprehension},
  author = {Levy, Omer and Seo, Minjoon and Choi, Eunsol and Zettlemoyer, Luke},
  booktitle = {Proceedings of the 21st Conference on Computational Natural Language Learning},
  year = {2017}
}
```