README.md

---
license: other
task_categories:
- text-generation
- fill-mask
tags:
- narrative infilling
- text infilling
- story understanding
- benchmark
- language model evaluation
language:
- en
pretty_name: NarrativeInfilling-Benchmark
size_categories:
- 1K<n<10K

rai:dataLimitations:
  - The dataset is English only and does not support multilingual evaluation
  - The benchmark spans four narrative domains; generalization to other genres or text types is not guaranteed
  - Missing spans are contiguous sentence sequences of one to three sentences; non-contiguous or stylistic gaps are not covered
  - Benchmark instances are drawn from publicly available corpora and may overlap with LLM pretraining data, though contamination analysis shows fewer than 0.15% of model responses exhibit near-verbatim reproduction

rai:dataBiases:
  - Source corpora reflect the demographic and cultural biases present in Wikipedia, CNN/DailyMail, ROCStories, and SIND
  - News articles (CNN/DailyMail) may contain temporal and political biases
  - Commonsense stories (ROCStories) reflect everyday scenarios predominantly from Western cultural contexts
  - Visual narratives (SIND) are grounded in Flickr image captions and may reflect photographer and platform biases

rai:personalSensitiveInformation:
  - All documents are drawn from publicly available sources
  - No individual personal identifiers are intentionally included
  - Users should screen for personally identifiable information before redistribution or downstream use

rai:dataUseCases:
  - Evaluating narrative infilling capabilities of instruction-tuned language models
  - Benchmarking automatic and qualitative evaluation metrics for text generation
  - Studying the effect of prompt design and reasoning guidance on narrative reconstruction
  - Use cases for which validity has not been established include open-domain generation, multilingual infilling, and non-narrative text reconstruction

prov:wasDerivedFrom:
  - name: Wikipedia
    url: https://huggingface.co/datasets/wikimedia/wikipedia
  - name: CNN/DailyMail
    url: https://huggingface.co/datasets/abisee/cnn_dailymail
  - name: ROCStories
    url: https://cs.rochester.edu/nlp/rocstories/
  - name: SIND (Sequential Image Narrative Dataset)
    url: https://visionandlanguage.net/VIST/dataset.html

prov:wasGeneratedBy:
  - name: Data Collection
    description: >
      Narratives were sampled from four publicly available datasets spanning 
      encyclopedic text (Wikipedia), news articles (CNN/DailyMail), commonsense 
      stories (ROCStories), and visual narratives (SIND). Narratives were filtered 
      to ensure a minimum length suitable for span masking.
  - name: Span Masking
    description: >
      One to three contiguous sentences were masked from each narrative to create 
      infilling instances. Blank positions (opening, middle, closing) were 
      systematically varied to ensure balanced coverage across narrative positions. 
      The masked span serves as the gold answer for evaluation.
  - name: Quality Filtering
    description: >
      Instances were filtered to remove duplicates and narratives where masking 
      produced degenerate or trivially short contexts.
---

# Dataset Card for NarrativeInfilling Benchmark

## Dataset Details

### Dataset Description

The Narrative Infilling Benchmark is a large-scale evaluation dataset for 
**narrative infilling** the task of generating a missing span within a 
narrative while maintaining consistency with both the preceding and following 
context. The benchmark spans four narrative domains and contains **9,142 instances** 
with systematic variation in blank position and span length, enabling controlled 
evaluation of language model infilling capabilities.

Each instance provides a narrative with one masked span (indicated by `____`) 
and the corresponding gold answer. The benchmark is designed to evaluate 
instruction-tuned LLMs under varying prompt specificity and reasoning guidance.

## Uses

### Direct Use
- **Narrative infilling evaluation:** Assess how well LLMs reconstruct missing 
narrative spans across diverse genres and blank positions.
- **Prompt sensitivity analysis:** Study the effect of instruction specificity 
and reasoning paradigms on Story Completion.
- **Metric evaluation:** Benchmark automatic metrics against qualitative 
human judgments for narrative infilling tasks.

### Out-of-Scope Use
- Not suitable for training language models directly without appropriate 
data splits to prevent leakage.
- Not suitable for factual question answering or non-narrative text reconstruction.

## Dataset Structure

**Format:** Single CSV file.

### Fields

| Column | Description |
|--------|-------------|
| `dataset` | Source domain: `wikipedia`, `cnn_dailymail`, `roc`, `sind` |
| `ref_id` | Integer reference ID for the instance (0–999 per dataset) |
| `source_text` | Full narrative text  |
| `problem` | The infilling problem presented to the model with masked span indicated by `____` |
| `gold_answer` | The original masked span serving as the reference answer |
| `n` | Number of sentences in the masked span (1, 2, or 3) |
| `unit_idx` | Position index of the masked span within the narrative |


### Splits

The dataset is distributed as a single file. Users can split by `dataset` 
column for domain-specific evaluation or by `unit_idx` for position-specific analysis.

## Dataset Creation

### Curation Rationale

No existing benchmark directly evaluates LLM narrative infilling across 
multiple genres with controlled variation in blank position and span length. 
This benchmark fills that gap by providing a standardized multi-domain 
evaluation setting for narrative reconstruction.

### Source Data

Instances were derived from four publicly available datasets:

| Domain | Source | Narrative Type |
|--------|--------|---------------|
| `wikipedia` | Wikipedia | Encyclopedic text |
| `cnn_dailymail` | CNN/DailyMail | News articles |
| `roc` | ROCStories | Commonsense stories |
| `sind` | SIND | Sequential Image Narrative Dataset |

### Annotation Process

No additional human annotation was performed during dataset construction. 
The gold answers are the original masked sentences from the source texts. 
Human evaluation of model outputs was conducted separately using a 
five-dimensional qualitative rubric (Fluency, Context Faithfulness, 
Bidirectional Coherence, Narrative Consistency, Informativeness) as 
described in the accompanying paper.

## Bias, Risks, and Limitations

- **Domain shift:** Performance varies substantially across domains; 
models strong on CNN/DailyMail may not generalize to SIND.
- **Pretraining overlap:** Source corpora are publicly available and 
may appear in LLM pretraining data. Contamination analysis in the 
accompanying paper shows a very small fraction of responses exhibit 
near-verbatim reproduction.
- **Contiguous spans only:** Real-world narrative gaps may involve 
non-contiguous or stylistic missing content not covered by this benchmark.

## Citation

```bibtex
@inproceedings{narrativeinfilling2025,
  title  = {Evaluating Narrative Infilling in Large Language Models},
  year   = {2025},
}
```

## More Information

This release supports the reproducibility of the results reported in the 
accompanying paper. The benchmark data, evaluation code, and model 
outputs are publicly available in the accompanying repository.