README.md

---
license: cc-by-sa-4.0
task_categories:
  - multiple-choice
  - question-answering
language:
  - ar
  - bn
  - de
  - en
  - es
  - fr
  - id
  - ja
  - pt
  - ru
  - sw
  - zh
configs:
  - config_name: ar
    data_files:
      - split: train
        path: data/ar/train.parquet
      - split: validation
        path: data/ar/validation.parquet
      - split: test
        path: data/ar/test.parquet
  - config_name: bn
    data_files:
      - split: train
        path: data/bn/train.parquet
      - split: validation
        path: data/bn/validation.parquet
      - split: test
        path: data/bn/test.parquet
  - config_name: de
    data_files:
      - split: train
        path: data/de/train.parquet
      - split: validation
        path: data/de/validation.parquet
      - split: test
        path: data/de/test.parquet
  - config_name: en
    data_files:
      - split: train
        path: data/en/train.parquet
      - split: validation
        path: data/en/validation.parquet
      - split: test
        path: data/en/test.parquet
  - config_name: es
    data_files:
      - split: train
        path: data/es/train.parquet
      - split: validation
        path: data/es/validation.parquet
      - split: test
        path: data/es/test.parquet
  - config_name: fr
    data_files:
      - split: train
        path: data/fr/train.parquet
      - split: validation
        path: data/fr/validation.parquet
      - split: test
        path: data/fr/test.parquet
  - config_name: id
    data_files:
      - split: train
        path: data/id/train.parquet
      - split: validation
        path: data/id/validation.parquet
      - split: test
        path: data/id/test.parquet
  - config_name: ja
    data_files:
      - split: train
        path: data/ja/train.parquet
      - split: validation
        path: data/ja/validation.parquet
      - split: test
        path: data/ja/test.parquet
  - config_name: pt
    data_files:
      - split: train
        path: data/pt/train.parquet
      - split: validation
        path: data/pt/validation.parquet
      - split: test
        path: data/pt/test.parquet
  - config_name: ru
    data_files:
      - split: train
        path: data/ru/train.parquet
      - split: validation
        path: data/ru/validation.parquet
      - split: test
        path: data/ru/test.parquet
  - config_name: sw
    data_files:
      - split: train
        path: data/sw/train.parquet
      - split: validation
        path: data/sw/validation.parquet
      - split: test
        path: data/sw/test.parquet
  - config_name: zh
    data_files:
      - split: train
        path: data/zh/train.parquet
      - split: validation
        path: data/zh/validation.parquet
      - split: test
        path: data/zh/test.parquet
  - config_name: parallel
    data_files:
      - split: train
        path: data/parallel/train.parquet
      - split: validation
        path: data/parallel/validation.parquet
      - split: test
        path: data/parallel/test.parquet
---

# PolyFact

PolyFact is a parallel multilingual multiple-choice QA dataset for evaluating and
post-training factual recall across 12 languages. Every item is grounded in a
Wikidata triple and aligned across Arabic, Bengali, German, English, Spanish,
French, Indonesian, Japanese, Portuguese, Russian, Swahili, and Chinese.

This curated release contains **58,807 facts**:

| Split | Facts per configuration |
|---|---:|
| Train | 56,324 |
| Validation | 444 |
| Test | 2,039 |


## Loading the dataset

PolyFact has no default configuration, so a configuration name is required.

```python
from datasets import load_dataset

# One language: ar, bn, de, en, es, fr, id, ja, pt, ru, sw, or zh
en = load_dataset("anonymous/PolyFact", "en")

# All 12 languages aligned by fact
parallel = load_dataset("anonymous/PolyFact", "parallel")
```

## Configurations and schema

| Configuration | Shape | Main fields |
|---|---|---|
| Language (`en`, `de`, etc.) | One row per fact and language | `fact_id`, `question`, `option_a`–`option_d`, `answer_text`, `answer_index`, `option_ids` |
| `parallel` | One row per fact | Wikidata subject/property/object IDs plus a `translations` struct containing all 12 languages |

Options are independently shuffled in each language. Use `option_ids` to align
candidate entities exactly across languages; do not assume option positions match.

## What was cleaned

The current release removes high-ambiguity or intrinsically multi-valued relations,
corrupt labels, defective or answer-leaking questions, foreign-script artifacts,
split overlap, duplicate options, and over-represented answers. Distractors were
resampled using Wikidata entity IDs, and the final test split was audited item by
item; 484 test facts were removed. The associated droplists and audit summaries are
included in this repository.

All 13 configurations contain the same fact IDs in each split. The release was
validated for answer/index/QID agreement, cross-language option-set alignment,
duplicate rows, and cross-split overlap.

## Limitations

Questions were initially generated independently per language and may still contain
subtle translation or entity-reference errors despite extensive automated and human
auditing. PolyFact inherits Wikidata's coverage and labeling biases, and the curated
set is weighted toward relations with a single unambiguous answer. Results measure
ranking among four candidates rather than unrestricted factual generation.