Datasets:

vigneshwar234
/

spiral-bench

question_id stringclasses 8 values	question stringclasses 8 values	question_type stringclasses 2 values	passage_id stringlengths 11 20	stance int64 -1 1	stance_label stringclasses 2 values	text stringlengths 123 175
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p00	1	supporting	Fleet study finds electric vehicles reduce fuel spending by 60 percent per kilometre, producing clear savings in total running cost over five years of ownership.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p01	1	supporting	Maintenance records show electric drivetrains need fewer serviced parts, so scheduled servicing cost falls sharply and owners save money across the ownership period.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p02	1	supporting	Home charging at off peak tariffs cuts energy cost per kilometre well below petrol, and the savings grow as annual distance driven increases.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p03	1	supporting	Government purchase incentives and road tax exemptions lower the effective purchase price, improving the total cost of ownership for electric cars in most markets.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p04	-1	opposing	Insurance premiums for electric vehicles run 20 to 40 percent higher because repair networks are scarce, which offsets a large share of the fuel savings.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p05	-1	opposing	Battery replacement outside warranty is a significant expense, and depreciation on older electric models has been steeper than on comparable petrol cars.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p06	-1	opposing	For low annual mileage drivers the higher purchase price is never recovered, so the petrol car remains cheaper across a realistic ownership period.
ev_cost	is an electric car cheaper to own than a petrol car	contested	ev_cost_p07	-1	opposing	Drivers without home charging rely on public rapid charging, where cost per kilowatt hour approaches petrol equivalence and erodes the expected running cost advantage.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p00	1	supporting	Supplementation corrects deficiency reliably and improves measured bone mineral density in adults with low baseline serum levels, supporting a daily supplement recommendation.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p01	1	supporting	Meta analysis of fracture prevention trials reports a modest benefit in older adults when vitamin D is combined with calcium over sustained daily use.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p02	1	supporting	Observational cohorts associate higher vitamin D status with better immune function and fewer respiratory infections across winter months in northern latitudes.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p03	1	supporting	Muscle strength and reported fall rates improve in deficient populations after daily supplementation, indicating a functional benefit beyond bone chemistry.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p04	-1	opposing	The VITAL randomised trial enrolled 25871 participants and found no significant effect of vitamin D supplementation on cardiovascular events or cancer incidence.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p05	-1	opposing	A Cochrane review reports no reduction in all cause mortality attributable to vitamin D supplementation in adults who are not deficient at baseline.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p06	-1	opposing	Routine supplementation in populations with adequate serum levels shows no measurable outcome improvement, and confers unnecessary cost and pill burden.
vitamin_d	should adults take vitamin D supplements daily	contested	vitamin_d_p07	-1	opposing	High dose supplementation has been linked to hypercalcaemia and, in some trials, to an increase rather than a decrease in fall risk among older adults.
remote_work	does remote work increase employee productivity	contested	remote_work_p00	1	supporting	A randomised call centre experiment measured a 13 percent increase in completed calls among remote staff, attributed to a quieter environment and fewer breaks.
remote_work	does remote work increase employee productivity	contested	remote_work_p01	1	supporting	Eliminating commuting returns usable hours to the working day, and self reported focus time rises substantially for roles built around individual deep work.
remote_work	does remote work increase employee productivity	contested	remote_work_p02	1	supporting	Attrition falls under flexible remote policies, and lower turnover reduces the productivity loss that accompanies repeated hiring and retraining cycles.
remote_work	does remote work increase employee productivity	contested	remote_work_p03	1	supporting	Output per engineer measured by completed tickets held steady or improved after a full transition to distributed work in several large technology firms.
remote_work	does remote work increase employee productivity	contested	remote_work_p04	-1	opposing	Studies of software teams report slower onboarding and weaker mentoring for junior staff under remote arrangements, depressing team level output over time.
remote_work	does remote work increase employee productivity	contested	remote_work_p05	-1	opposing	Collaboration network analysis shows communication becomes more siloed after a remote transition, with fewer cross team connections and slower information flow.
remote_work	does remote work increase employee productivity	contested	remote_work_p06	-1	opposing	A large field experiment found remote data entry workers were 18 percent less productive than their office based counterparts on identical task sets.
remote_work	does remote work increase employee productivity	contested	remote_work_p07	-1	opposing	Self reported productivity gains frequently fail to appear in objective output measures, indicating a measurement gap rather than a genuine improvement.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p00	1	supporting	Standard competitive labour market models predict that a binding wage floor above equilibrium reduces quantity of labour demanded, producing measurable job losses.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p01	1	supporting	Studies of large local increases report reduced scheduled hours for low wage workers, so total earnings gains are smaller than the headline wage rise suggests.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p02	1	supporting	Employers in low margin sectors respond to mandated wage increases by accelerating automation, reducing the number of entry level positions available.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p03	1	supporting	Teen and low skill employment rates fell in several regions following sharp minimum wage increases relative to comparable neighbouring regions.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p04	-1	opposing	The Card and Krueger fast food study compared neighbouring states and found no employment decline following a minimum wage increase in New Jersey.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p05	-1	opposing	Monopsony models show that where employers hold wage setting power, a moderate wage floor can raise both wages and employment simultaneously.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p06	-1	opposing	Meta analysis across many studies reports employment effects clustered near zero, with publication bias inflating the appearance of negative findings.
min_wage	does raising the minimum wage reduce employment	contested	min_wage_p07	-1	opposing	Reduced turnover and improved worker retention after wage increases offset labour cost rises, leaving headcount broadly unchanged in observed firms.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p00	1	supporting	Time restricted eating reduces total daily energy intake for many participants without deliberate calorie counting, producing weight loss over twelve week trials.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p01	1	supporting	Adherence is higher for some participants under a simple eating window rule than under continuous calorie restriction, improving sustained weight outcomes.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p02	1	supporting	Fasting periods improved measured insulin sensitivity in several trials, an effect associated with improved fat oxidation during the fasted state.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p03	1	supporting	Participants following an eating window reported reduced evening snacking, a behaviour change that accounts for a meaningful share of the energy deficit.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p04	-1	opposing	Head to head trials matching total calories find no weight loss advantage for intermittent fasting over continuous restriction across twelve months.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p05	-1	opposing	A randomised trial of time restricted eating reported greater loss of lean mass in the fasting arm, an unfavourable body composition outcome.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p06	-1	opposing	Dropout rates in fasting arms exceed those in standard restriction arms in several studies, undermining claims of superior long term adherence.
fasting	is intermittent fasting effective for weight loss	contested	fasting_p07	-1	opposing	Benefits reported in short trials largely disappear once total energy intake is controlled, indicating the deficit rather than meal timing drives results.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p00	1	supporting	Lifecycle emissions per unit of nuclear electricity are comparable to wind and lower than solar, making it one of the lowest carbon generation sources available.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p01	1	supporting	Nuclear plants supply firm dispatchable output independent of weather, reducing the storage capacity a decarbonised grid would otherwise require.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p02	1	supporting	Countries with large existing nuclear fleets achieved grid decarbonisation faster than comparable countries relying on renewables alone over the same period.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p03	1	supporting	Land use per unit of energy delivered is far lower for nuclear than for wind or solar, easing siting constraints in densely populated regions.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p04	-1	opposing	Recent reactor projects in Europe and North America ran years late and multiples over budget, making levelised cost uncompetitive with renewables.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p05	-1	opposing	Construction timelines of a decade or more mean new nuclear cannot deliver emissions reductions inside the window that near term climate targets require.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p06	-1	opposing	No country has yet commissioned a permanent high level waste repository at scale, leaving long term disposal costs and liabilities unresolved.
nuclear	is nuclear power a good solution for reducing carbon emissions	contested	nuclear_p07	-1	opposing	Capital intensity concentrates financial risk in single large projects, and several utilities have faced severe distress after cost overruns on reactors.
smoking	does smoking cause lung cancer	settled	smoking_p00	1	supporting	Cohort studies following millions of participants show lung cancer incidence rising steeply with cumulative cigarette exposure in a clear dose response relationship.
smoking	does smoking cause lung cancer	settled	smoking_p01	1	supporting	Tobacco smoke contains identified carcinogens that form DNA adducts in bronchial epithelium, providing a direct biological mechanism for tumour initiation.
smoking	does smoking cause lung cancer	settled	smoking_p02	1	supporting	Lung cancer rates fall measurably in populations after sustained declines in smoking prevalence, consistent with a causal rather than correlational link.
smoking	does smoking cause lung cancer	settled	smoking_p03	1	supporting	Risk declines progressively after cessation, and the size of the decline tracks years since quitting, supporting causation over confounding.
smoking	does smoking cause lung cancer	settled	smoking_p04	1	supporting	Case control studies conducted across many countries and decades report consistent and large effect sizes linking smoking to lung cancer.
smoking	does smoking cause lung cancer	settled	smoking_p05	1	supporting	Animal exposure experiments reproduce airway tumours, corroborating the epidemiological association with controlled experimental evidence.
smoking	does smoking cause lung cancer	settled	smoking_p06	1	supporting	Major public health authorities classify tobacco smoking as an established cause of lung cancer based on the full weight of accumulated evidence.
smoking	does smoking cause lung cancer	settled	smoking_p07	1	supporting	Second hand smoke exposure raises lung cancer risk in non smokers, extending the causal relationship beyond direct smoking behaviour.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p00	1	supporting	Two doses of measles containing vaccine confer approximately 97 percent protection against infection in controlled effectiveness studies.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p01	1	supporting	National measles incidence fell by more than 99 percent following the introduction of routine childhood vaccination programmes.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p02	1	supporting	Outbreak investigations consistently find attack rates concentrated among unvaccinated individuals within otherwise well covered communities.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p03	1	supporting	Serological studies confirm durable neutralising antibody titres persisting for decades after the standard two dose vaccination schedule.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p04	1	supporting	Regions achieving sustained high coverage have interrupted endemic transmission entirely, consistent with strong vaccine mediated population protection.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p05	1	supporting	Measles mortality declined sharply worldwide in step with expanding vaccination coverage across low and middle income countries.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p06	1	supporting	Post exposure vaccination within 72 hours reduces the probability of clinical disease, demonstrating direct protective action of the vaccine.
vaccines_measles	do measles vaccines prevent measles infection	settled	vaccines_measles_p07	1	supporting	Resurgence of cases follows measurable declines in coverage, confirming the protective effect operates at population level.

SPIRAL-Bench v0

A small benchmark for testing whether the wording of a retrieval query changes the balance of the evidence a retriever returns.

Built for the Ouroboros project, which studies self-confirming retrieval loops in agentic RAG.

The question this dataset exists to answer

In agentic RAG, the system writes its own follow-up search queries, and it writes them using what it currently believes. A retriever ranks documents by how closely they match the wording of the query it receives. If those two facts combine, an agent can retrieve progressively more one-sided evidence without any document being false and without any attacker involved.

Testing that claim needs a corpus where balance is guaranteed rather than assumed. That is what this dataset provides.

What is in it

Three configs.

`passages` (64 rows)

Every contested question carries exactly four supporting and four opposing passages, so the evidence is balanced by construction. Settled questions are deliberately one-sided and act as a control.

field	description
`question_id`	short identifier, for example `ev_cost`
`question`	the neutral question a user would ask
`question_type`	`contested` or `settled`
`passage_id`	unique passage identifier
`stance`	`1` supporting, `-1` opposing
`stance_label`	`supporting` or `opposing`
`text`	the passage

`queries` (48 rows)

Two query sequences per question, three hops each, for a controlled comparison.

arm	what it does
`drift`	wording bends toward the belief, for example "is an electric car cheaper than petrol" becomes "savings from owning an electric car"
`neutral`	wording stays a neutral restatement across all three hops

Everything else is held fixed between the arms, so any difference in retrieved balance is attributable to wording alone.

`traces` (48 rows)

The logged output of the first experiment: one row per retrieval, carrying the query, the drift score, the support and oppose counts, and the retrieved passages.

Questions covered

Contested (balanced 4 / 4): electric vehicle ownership cost, vitamin D supplementation, remote work productivity, minimum wage and employment, intermittent fasting, nuclear power.

Settled (one-sided control): smoking and lung cancer, measles vaccine effectiveness.

Result obtained on this dataset

Handing the retriever an oracle query, built only from vocabulary unique to one side, retrieves 100% supporting evidence. The reverse oracle retrieves 100% opposing evidence.

So query wording alone moves retrieved evidence across a range of 2.00, on a corpus that is verifiably balanced. Supporting and opposing passages share only 12.9% of their vocabulary, so stance is separable in principle.

Hand-written drift queries captured only 4.2% of that available range, and a drift metric computed over a TF-IDF encoder stayed flat. The mechanism is available; the first probe was too weak to trigger it.

Honest limitations

Read these before using the dataset.

The passages are synthetic. They are research-style text written for this prototype, not excerpts from real papers. Claims inside them describe real findings in outline, but they are not quotations and must not be cited as literature.
It is small. 8 questions and 64 passages. Enough to test a mechanism, not enough for statistical claims about model behaviour.
Queries in v0 are authored, not agent-generated. They are deliberate probes written to represent belief-aligned drift. Whether a live language model writes queries like these is an open question this dataset does not answer.
Balance is balanced by construction, not by sampling. Real corpora are not balanced, and results here should not be read as an estimate of real-world evidence distributions.

Loading

from datasets import load_dataset

passages = load_dataset("vigneshwar234/spiral-bench", "passages", split="train")
queries  = load_dataset("vigneshwar234/spiral-bench", "queries",  split="train")
traces   = load_dataset("vigneshwar234/spiral-bench", "traces",   split="train")

Reproducing the experiment

Code lives in the Ouroboros repository. It needs numpy, scikit-learn and matplotlib, and runs on a laptop CPU with no GPU, no API key and no model download.

python3 run_experiment.py
python3 run_diagnostic.py

Planned for v1

A larger question set, a dense sentence encoder in place of TF-IDF, and query sequences written by a live language model rather than by hand.

Citation

@misc{spiralbench2026,
  title  = {SPIRAL-Bench: a balanced-evidence benchmark for retrieval query bias},
  author = {Vigneshwar L},
  year   = {2026},
  note   = {Ouroboros project, Department of Artificial Intelligence and Data Science}
}