The SMeL Test

Data for The SMeL Test (Simple Media Literacy Test), a benchmark that measures whether language models appropriately weight information by the trustworthiness of its source. Every document in the dataset is a short passage about a fictional entity (a government agency, a crime, or a natural disaster) attributed to one of seven sources spanning a spectrum of reliability, from encyclopedias and newspapers to anonymous forum posts and fan fiction. Because the entities are invented, a model cannot rely on prior knowledge and must reason from the in-context documents and their sources alone.

• Paper: coming soon
• Code / usage examples: github.com/gahdritz/smel
• License: CC BY 4.0

Sources and trust labels

Passages are attributed to the following sources. The benchmark treats the first group as trustworthy and the second as untrustworthy:

Trustworthy	Untrustworthy
https://wikipedia.org, https://nytimes.com, https://britannica.com	https://reddit.com, https://4chan.com, https://fanfiction.net, unknown (an anonymous, conspiratorial manifesto)

Configurations

The dataset ships four loadable configurations.

Config	Rows	Description
qa_ground_truth	600	Seed documents (one per entity) from which every source-specific variant was generated. Source-agnostic.
qa_rewritten	4,200	Each seed rewritten in the voice/style of each of the 7 sources, preserving the original answer.
qa_corrupted	4,200	Source-specific variants whose passage supports a different (incorrect) answer to the entity's question.
summarization	4,200	Single-fact, source-attributed passages used for the "active filtering" summarization task.

4,200 = 600 entities × 7 sources.

QA fields (qa_rewritten, qa_corrupted)

Field	Type	Description
source	string	The source the passage is attributed to (see table above).
passage	string	The document text shown to the model in context.
question	string	A question about the entity that the passage bears on.
answer	string	The answer the passage supports — the correct answer in rewritten, an incorrect one in corrupted.
entity	string	Name of the fictional entity the passage is about (e.g. "American Harvest Preservation Agency").
entity_type	string	One of agency, crime, disaster.

Seed fields (qa_ground_truth)

The seed documents are source-agnostic, so this config has the same fields as above minus source: passage, question, answer, entity, entity_type.

Summarization fields (summarization)

Field	Type	Description
source	string	The source the passage is attributed to.
entity	string	Name of the fictional entity.
entity_type	string	One of agency, crime, disaster.
fact	string	The single atomic fact the passage conveys.
passage	string	The source-styled document text.

Example usage

from datasets import load_dataset

# Rewritten (trustworthy-through-untrustworthy) QA variants
rewritten = load_dataset("gahdritz/smel", name="qa_rewritten", split="test")

# Corrupted variants (same questions, wrong answers)
corrupted = load_dataset("gahdritz/smel", name="qa_corrupted", split="test")

# Seed documents the variants were generated from
seeds = load_dataset("gahdritz/smel", name="qa_ground_truth", split="test")

# Summarization / active-filtering documents
summ = load_dataset("gahdritz/smel", name="summarization", split="test")

print(rewritten[0])

To reconstruct a task instance, group by entity and select passages from the desired combination of sources. See the GitHub repository for the full evaluation harness.

Dataset creation

All passages are synthetic. Fictional entities (agencies, crimes, disasters) and their seed documents were generated with large language models, along with a question/answer pair per entity. Each seed was then (a) rewritten in the characteristic style of each source to produce qa_rewritten, and (b) altered to support an alternative answer to produce qa_corrupted. The summarization split was generated analogously from single atomic facts.

Responsible AI

Bias

The benchmark encodes an explicit, coarse prior about source reliability: it labels Wikipedia, The New York Times, and Encyclopedia Britannica as trustworthy and Reddit, 4chan, fan fiction, and anonymous manifestos as untrustworthy. This binary is a deliberate simplification — real-world reliability is contextual, and reputable outlets can err while informal sources can be correct. Treating these labels as universal ground truth would itself be a bias. In addition, the entities, place names, and events are predominantly US- and Western-centric, the data is English-only, and, because the passages are LLM-generated, they inherit stylistic and topical biases of the generating models.

Limitations

The data is entirely synthetic and does not capture the full complexity of real-world misinformation, adversarial sourcing, or subtle credibility cues. The set of sources is small and fixed, trust is modeled as a binary rather than a graded or context-dependent property, and each instance isolates a single question or fact. "Correct" behavior on the benchmark reflects the authors' stipulated trust priors and should not be read as a claim about the reliability of any specific real platform.

Social impact

The dataset is intended to support research into media literacy and model robustness to unreliable sources — an increasingly important property as models are used to synthesize information from heterogeneous, uncurated inputs. Potential negative uses include over-generalizing its stipulated trust labels to real sources, or citing benchmark results to justify blanket distrust (or trust) of particular platforms. Because all entities are fictional, the dataset poses no privacy risk to real individuals and contains no real-world claims that could be mistaken for fact.

Citation

@misc{gahdritz_smel_2025,
  author    = {Gustaf Ahdritz},
  title     = {The SMeL Test},
  year      = {2025},
  publisher = {Hugging Face Hub},
  url       = {https://huggingface.co/datasets/gahdritz/smel}
}