jvonrad commited on
Commit
348f8a1
·
verified ·
1 Parent(s): 6c72209

De-anonymize loading example; add EACL 2027 paper link and BibTeX citation

Browse files
Files changed (1) hide show
  1. README.md +35 -2
README.md CHANGED
@@ -130,6 +130,8 @@ post-training factual recall across 12 languages. Every item is grounded in a
130
  Wikidata triple and aligned across Arabic, Bengali, German, English, Spanish,
131
  French, Indonesian, Japanese, Portuguese, Russian, Swahili, and Chinese.
132
 
 
 
133
  This curated release contains **58,807 facts**:
134
 
135
  | Split | Facts per configuration |
@@ -147,10 +149,10 @@ PolyFact has no default configuration, so a configuration name is required.
147
  from datasets import load_dataset
148
 
149
  # One language: ar, bn, de, en, es, fr, id, ja, pt, ru, sw, or zh
150
- en = load_dataset("anonymous/PolyFact", "en")
151
 
152
  # All 12 languages aligned by fact
153
- parallel = load_dataset("anonymous/PolyFact", "parallel")
154
  ```
155
 
156
  ## Configurations and schema
@@ -184,3 +186,34 @@ auditing. PolyFact inherits Wikidata's coverage and labeling biases, and the cur
184
  set is weighted toward relations with a single unambiguous answer. Results measure
185
  ranking among four candidates rather than unrestricted factual generation.
186
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
130
  Wikidata triple and aligned across Arabic, Bengali, German, English, Spanish,
131
  French, Indonesian, Japanese, Portuguese, Russian, Swahili, and Chinese.
132
 
133
+ **Paper:** [PolyFact: Comparing Consistency-Driven Post-training Methods for Cross-Lingual Factual Recall](https://arxiv.org/abs/2606.06586), **EACL 2027 (Main Conference)**. Please cite it if you use the dataset (see [Citation](#citation)).
134
+
135
  This curated release contains **58,807 facts**:
136
 
137
  | Split | Facts per configuration |
 
149
  from datasets import load_dataset
150
 
151
  # One language: ar, bn, de, en, es, fr, id, ja, pt, ru, sw, or zh
152
+ en = load_dataset("jvonrad/PolyFact", "en")
153
 
154
  # All 12 languages aligned by fact
155
+ parallel = load_dataset("jvonrad/PolyFact", "parallel")
156
  ```
157
 
158
  ## Configurations and schema
 
186
  set is weighted toward relations with a single unambiguous answer. Results measure
187
  ranking among four candidates rather than unrestricted factual generation.
188
 
189
+ ## Citation
190
+
191
+ PolyFact was introduced in the following paper, accepted to the **EACL 2027 Main Conference**.
192
+ If you use the dataset, please cite:
193
+
194
+ ```bibtex
195
+ @inproceedings{vonrad2027polyfact,
196
+ title = {PolyFact: Comparing Consistency-Driven Post-training Methods for Cross-Lingual Factual Recall},
197
+ author = {Jonathan von Rad and Louis Arts and George Burgess and Eleftheria Kolokytha and Harry O'Donnell and Ektor Oikonomidis Doumpas and Eduardo Sanchez and Yao Lu and Pontus Stenetorp},
198
+ booktitle = {Proceedings of the 2027 Conference of the European Chapter of the Association for Computational Linguistics (EACL 2027), Main Conference},
199
+ year = {2027},
200
+ eprint = {2606.06586},
201
+ archivePrefix = {arXiv},
202
+ primaryClass = {cs.CL},
203
+ url = {https://arxiv.org/abs/2606.06586}
204
+ }
205
+ ```
206
+
207
+ arXiv preprint version:
208
+
209
+ ```bibtex
210
+ @misc{vonrad2026polyfactcomparingconsistencydrivenposttraining,
211
+ title={PolyFact: Comparing Consistency-Driven Post-training Methods for Cross-Lingual Factual Recall},
212
+ author={Jonathan von Rad and Louis Arts and George Burgess and Eleftheria Kolokytha and Harry O'Donnell and Ektor Oikonomidis Doumpas and Eduardo Sanchez and Yao Lu and Pontus Stenetorp},
213
+ year={2026},
214
+ eprint={2606.06586},
215
+ archivePrefix={arXiv},
216
+ primaryClass={cs.CL},
217
+ url={https://arxiv.org/abs/2606.06586},
218
+ }
219
+ ```