README.md

---
license: mit
task_categories:
  - question-answering
  - text-retrieval
language:
  - en
tags:
  - temporal-qa
  - time-sensitive
  - retrieval-augmented-generation
  - timestamp
  - streamingqa
  - timeqa
size_categories:
  - 10K<n<100K
---

# Time-Sensitive QA Dataset (StreamingQA + TimeQA)

A time-sensitive question answering dataset combining **StreamingQA** and **TimeQA** for evaluating temporal reasoning in RAG systems.

## Dataset Description

This dataset contains question-answer pairs where the correct answer depends on a specific timestamp.
Each question is prefixed with a date (e.g., "Today is Tuesday, September 24, 2013.") and the model
must use temporal context to provide the correct answer for that point in time.

### Sources

- **StreamingQA** ([Liska et al., 2022](https://arxiv.org/abs/2205.02266)): Time-sensitive questions about news articles from the WMT corpus
- **TimeQA** ([Chen et al., 2021](https://arxiv.org/abs/2108.06314)): Temporal knowledge base questions derived from WikiData

### Features

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique identifier |
| `question` | string | Full question with timestamp prefix (e.g., "Today is Monday, Jan 5, 2015. Who is...") |
| `context` | string | Evidence document containing the answer |
| `answers` | list[string] | List of acceptable answers |
| `source` | string | Original dataset (`streamingqa` or `timeqa`) |
| `question_timestamp` | string | When the question is asked (ISO format) |
| `evidence_timestamp` | string | When the evidence document was published |
| `dimensions` | string | Metadata about temporal dimensions (JSON) |
| `is_hard_negative` | bool | Whether this is a hard negative (temporal perturbation) |
| `original_question` | string | Question without timestamp prefix |

### Splits

| Split | Examples | Description |
|-------|----------|-------------|
| train | 25,041 | Training set |
| validation | 2,800 | Validation set |
| test | 9,089 | Test/evaluation set |

### Hard Negatives

Examples with `is_hard_negative=True` are temporal perturbations where the question timestamp
is shifted to make the context outdated or incorrect. These are useful for evaluating a model's
ability to detect temporal mismatches between query time and evidence time.

## Usage

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Catkamakura/ts-qa")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# Example
example = train[0]
print(example["question"])
# "Today is Tuesday, September 24, 2013. Who was the Transport Minister of Thailand?"
print(example["answers"])
# ["Chadchart Sittipunt"]
print(example["source"])
# "streamingqa"

# Filter by source
streamingqa_examples = train.filter(lambda x: x["source"] == "streamingqa")
timeqa_examples = train.filter(lambda x: x["source"] == "timeqa")

# Filter hard negatives
hard_negatives = train.filter(lambda x: x["is_hard_negative"])
```

## Citation

If you use this dataset, please cite the original sources:

```bibtex
@inproceedings{liska2022streamingqa,
  title={StreamingQA: A Benchmark for Adaptation to New Knowledge over Time in Question Answering Models},
  author={Liska, Adam and Kocisky, Tomas and Gribovskaya, Elena and Terber, Tayfun and Shutova, Ekaterina and Glaese, Amelia},
  booktitle={International Conference on Machine Learning},
  pages={13604--13622},
  year={2022},
  organization={PMLR}
}

@inproceedings{chen2021timeqa,
  title={A Dataset for Answering Time-Sensitive Questions},
  author={Chen, Wenhu and Wang, Xinyi and Wang, William Yang},
  booktitle={NeurIPS 2021 Datasets and Benchmarks Track},
  year={2021}
}
```

## License

MIT License