driftlense1/driftlense

DriftLens: Questions

This dataset accompanies the DriftLens framework for measuring reasoning robustness in large language models on unverifiable, open-ended questions — prompts where there is no single objective ground truth, only a plausible chain of reasoning.

Construction

We construct a benchmark of questions that are:

• Reasoning-invoking — requiring trade-offs, comparisons, or justification;
• Unverifiable — admitting no single objective ground truth;
• Persona-indifferent — answerable independently of the asker's demographic attributes;
• Stand-alone — comprehensible without hidden context.

Candidates are aggregated from eight public sources spanning career, ethics, finance, legal, medical, and daily-dilemma questions:

Source	Domain	Reference
Career Guidance	Career	career_guidance_dataset
ETHICS (Commonsense, Justice)	Ethics	Hendrycks et al., 2021
Medical QA	Health	malikeh1375/medical-question-answering-datasets, 2025
Financial Advisor-100	Finance	nihiluis/financial-advisor-100, 2025
Legal Advice	Legal	Li et al., 2022
1M Reddit Questions	Open-domain	socialgrep/one-million-reddit-questions, 2025
DailyDilemmas	Daily ethics	Chiu et al., 2024

Candidates pass through a two-stage filter. First, an LLM rater (GPT-OSS-120B) applies the four criteria above to discard purely informational, jurisdiction-dependent, or persona-specific items. Second, three human annotators per question rate:

1. clarity,
2. freedom from vague or subjective terms,
3. persona-orthogonality across four dimensions (culture, age, gender, education),
4. absence of a universal ground truth,
5. reasoning-invoking nature.

Two releases are produced from this pipeline, differing only in the human-agreement threshold:

Config	Rows	Agreement
filtered	422	Unanimous (3/3) agreement on all checks. Used for the main experiments.
raw	1,061	Relaxed: 2/3 agreement.

The two files are independent passes of the same pipeline at two thresholds, not a nested subset. The filtered set is the recommended starting point; raw is provided for researchers who want broader coverage and can tolerate lower annotator agreement.

Data Fields

Field	Type	Description
question_index	int	Per-file row identifier. Indices are local to each file and are not stable across raw and filtered; do not use them to join the two files.
question	string	The open-ended question text (English).
label	string	Acceptance label from the curation pipeline. All released rows are ACCEPT.
reason	string	Short free-form justification for why the question qualifies as reasoning-dependent and unverifiable.
domains	string (list of strings)	One or more topical tags. Stored as a stringified Python list; parse with ast.literal_eval.

Each question carries 1–5 domain tags (mean ≈ 2.5); see the accompanying paper for the domain ontology and distribution figure.

Intended Use

These questions are intended as inputs to a reasoning-robustness evaluation pipeline. They are deliberately not paired with reference answers, because the dataset targets the unverifiable-task regime where reference answers do not exist or are not meaningful.

Typical use:

1. Sample a question q from the dataset.
2. Generate a baseline response r_0 from the model under test.
3. Generate one or more perturbed responses r_i under task-irrelevant context (e.g., user-attribute memory, persona, stylistic preface).
4. Compute a drift score between r_0 and r_i (e.g., DTW alignment cost over reasoning traces, or a Symbolic Recurrence Index combining Levenshtein and Jensen–Shannon distances over symbolic representations).

The dataset is not designed for:

• Training models toward correct answers (there are no answers).
• Benchmarks that require ground-truth labels (e.g., accuracy, F1).
• Use as a bias / fairness benchmark on its own — bias-style perturbations are one application, not a property of the questions themselves.

Limitations and Considerations

• English only. Questions are in English and reflect predominantly U.S./Anglophone framings of everyday dilemmas. Cross-cultural generalization should be assumed limited.
• No reference answers. Intentional, but means the dataset cannot be used directly for accuracy-style evaluation.
• Sensitive topics. Some questions touch on health, infidelity, family conflict, mortality, and contested ethical issues.
• Domain tags are author-assigned, not crowd-validated. They support stratified sampling and reporting, not precise multi-label classification.
• Source licensing. The eight upstream sources retain their own licenses; users should consult the original references before redistribution beyond research use.

Loading

from datasets import load_dataset

# Strict release (unanimous human agreement; recommended)
ds = load_dataset("SymbolicDrift/<repo-name>", "filtered", split="train")

# Relaxed release (2/3 human agreement)
ds_raw = load_dataset("SymbolicDrift/<repo-name>", "raw", split="train")

To parse the domains field:

import ast
ds = ds.map(lambda x: {"domains": ast.literal_eval(x["domains"])})

Citation

If you use this dataset, please cite the accompanying paper.

License

Released under CC BY 4.0 for the curation and annotations. Upstream source data is governed by the licenses of the original datasets cited above.