README.md

---
license: mit
language:
- en
size_categories:
- 1K<n<10K
configs:
  - config_name: ambiguous_story_task
    data_files:
      - split: test
        path: ambiguous_story_task.parquet
  - config_name: completion_of_failed_actions
    data_files:
      - split: test
        path: completion_of_failed_actions.parquet
  - config_name: discrepant_desires
    data_files:
      - split: test
        path: discrepant_desires.parquet
  - config_name: discrepant_emotions
    data_files:
      - split: test
        path: discrepant_emotions.parquet
  - config_name: discrepant_intentions
    data_files:
      - split: test
        path: discrepant_intentions.parquet
  - config_name: emotion_regulation
    data_files:
      - split: test
        path: emotion_regulation.parquet
  - config_name: false_belief_task
    data_files:
      - split: test
        path: false_belief_task.parquet
  - config_name: faux_pas_recognition_test
    data_files:
      - split: test
        path: faux_pas_recognition_test.parquet
  - config_name: hidden_emotions
    data_files:
      - split: test
        path: hidden_emotions.parquet
  - config_name: hinting_task_test
    data_files:
      - split: test
        path: hinting_task_test.parquet
  - config_name: knowledge_attention_links
    data_files:
      - split: test
        path: knowledge_attention_links.parquet
  - config_name: knowledge_pretend_play_links
    data_files:
      - split: test
        path: knowledge_pretend_play_links.parquet
  - config_name: moral_emotions
    data_files:
      - split: test
        path: moral_emotions.parquet
  - config_name: multiple_desires
    data_files:
      - split: test
        path: multiple_desires.parquet
  - config_name: percepts_knowledge_links
    data_files:
      - split: test
        path: percepts_knowledge_links.parquet
  - config_name: persuasion_story_task
    data_files:
      - split: test
        path: persuasion_story_task.parquet
  - config_name: prediction_of_actions
    data_files:
      - split: test
        path: prediction_of_actions.parquet
  - config_name: scalar_implicature_test
    data_files:
      - split: test
        path: scalar_implicature_test.parquet
  - config_name: strange_story_task
    data_files:
      - split: test
        path: strange_story_task.parquet
  - config_name: unexpected_outcome_test
    data_files:
      - split: test
        path: unexpected_outcome_test.parquet
---

# ToMBench (English-only, OLMES-ready mirror)

This is a clean, English-only mirror of **ToMBench** (Chen et al., ACL 2024).
Source: https://github.com/zhchen18/ToMBench  Paper: arXiv:2402.15052

## Why this mirror exists

The official ToMBench distribution is JSONL on GitHub with bilingual (Chinese/English)
fields and inconsistent type inference (some rows have option fields as numeric, others
as string — breaks `datasets.load_dataset("json", ...)`). This mirror:

- Drops all Chinese-language columns (`能力`, `故事`, `问题`, `选项A-D`)
- Coerces all option fields to string (no type ambiguity)
- Normalizes the answer key to single uppercase A/B/C/D (one row had `"A. "`)
- Adds a precomputed `gold_index` column (0..3) for direct use with `MultipleChoiceTask`
- Splits 20 subtasks into 20 separate configs (one Parquet per subtask)

## Schema

| Column | Type | Description |
|---|---|---|
| `id` | int | Source row index |
| `ability` | string | Cognitive ability being tested (e.g., "Belief: Location false beliefs") |
| `story` | string | Story narrative (English) |
| `question` | string | Question (English) |
| `option_a` ... `option_d` | string | The four answer options |
| `answer_key` | string | Correct answer letter, one of "A"/"B"/"C"/"D" |
| `gold_index` | int | Correct answer 0-based index (0-3) |

## Configs and sizes

| Config | Examples |
|---|---:|
| `ambiguous_story_task` | 200 |
| `completion_of_failed_actions` | 20 |
| `discrepant_desires` | 20 |
| `discrepant_emotions` | 40 |
| `discrepant_intentions` | 40 |
| `emotion_regulation` | 20 |
| `false_belief_task` | 600 |
| `faux_pas_recognition_test` | 560 |
| `hidden_emotions` | 80 |
| `hinting_task_test` | 103 |
| `knowledge_attention_links` | 20 |
| `knowledge_pretend_play_links` | 30 |
| `moral_emotions` | 40 |
| `multiple_desires` | 20 |
| `percepts_knowledge_links` | 40 |
| `persuasion_story_task` | 100 |
| `prediction_of_actions` | 20 |
| `scalar_implicature_test` | 200 |
| `strange_story_task` | 407 |
| `unexpected_outcome_test` | 300 |
| **Total** | **2860** |

## Loading

```python
from datasets import load_dataset
ds = load_dataset("HCAI-Lab/tombench-en", "false_belief_task", split="test")
print(ds[0])
```

## License

MIT (mirroring the upstream ToMBench license).

**For evaluation only.** ToMBench authors explicitly forbid using this dataset for training.

## Citation

```bibtex
@inproceedings{chen-etal-2024-tombench,
    title={ToMBench: Benchmarking Theory of Mind in Large Language Models},
    author={Chen, Zhuang and others},
    booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics},
    year={2024},
}
```

This mirror was prepared 2026-05-23 by glennmatlin (HCAI-Lab) for use in the OLMES
evaluation pipeline; see `eilab-gt/social-data-attribution`, branch `tom-evals-rebuttal`.