Hravan/semantic-sensitivity

Semantic Sensitivity Benchmark

A benchmark of 10,843 equivalence groups designed to measure how consistently large language models answer semantically equivalent questions phrased in different surface forms.

Each group contains 2–5 restatements of the same question with the same correct answer. The benchmark spans 39 question categories across factual retrieval and logical reasoning.

The dataset is procedurally generated by scripts/generate_ss_groups.py (seed 42) and is fully reproducible from the accompanying code release.

---

Dataset structure

Each entry in ss_groups.json is a JSON object with the following fields:

Field	Type	Description
id	int	Unique group identifier
category	str	Question category (see below)
answer_type	str	"yes_no" or "word"
questions	list[str]	2–5 phrasings of the same question
expected	str	Ground-truth answer

Example

{
  "id": 0,
  "category": "capital_word_order",
  "answer_type": "yes_no",
  "expected": "yes",
  "questions": [
    "Is Tirana the capital of Albania?",
    "Is the capital of Albania Tirana?",
    "Does Albania have Tirana as its capital?"
  ]
}

---

Categories

Factual

Category	Groups	Description
capital_word_order	200	Capital–country word-order variants
capital_retrieval	100	Open-ended capital retrieval
active_passive	100	Active vs. passive voice restatements
contrastive_negation	100	Contrastive negation restatements
geographic_containment	100	City/country containment questions
chemical_formula	100	Chemical formula identification
classification	100	Taxonomy / type-of questions
element_symbol	50	Periodic-table symbol retrieval
country_language	79	Official language of a country
country_currency	71	Currency of a country
continent	85	Continent of a country
largest_city	61	Largest city of a country

Logical

Category	Groups	Description
arithmetic_order	400	Comparison of two numbers (word-order variants)
arithmetic_large	300	Large-number arithmetic (> 10 000)
arithmetic_xlarge	300	Extra-large-number arithmetic
arithmetic_result	100	Basic arithmetic result queries
arithmetic_convoluted	225	Multi-step arithmetic with surface variation
large_arithmetic_result	100	Large-number result queries
multiplication_order	300	Multiplication with operand-order variants
multiplication_result	100	Multiplication result queries
subtraction_equivalence	225	Subtraction phrasing equivalences
comparison_symmetric	300	Symmetric comparison statements
comparison_convoluted	225	Complex comparison restatements
unit_equivalence	100	Unit-conversion equivalences
double_negation	200	Double-negation elimination
negation_arithmetic	150	Negation embedded in arithmetic
negation_depth	100	Nested negation (mixed depths)
negation_even	100	Nested negation — even depths only
negation_odd	100	Nested negation — odd depths only
negation_depth_0	840	Nested negation — depth 0
negation_depth_1	840	Nested negation — depth 1
negation_depth_2	840	Nested negation — depth 2
negation_depth_3	840	Nested negation — depth 3
negation_depth_4	840	Nested negation — depth 4
negation_depth_5	840	Nested negation — depth 5
negation_depth_6	840	Nested negation — depth 6
contrapositive	200	Contrapositive equivalences
de_morgan	200	De Morgan's law restatements
quantifier_scope	92	Quantifier-scope variants

---

Fine-tuning experiments

The negation_depth_0–negation_depth_6 categories (840 groups each) are used in the depth-anchoring fine-tuning experiments described in the accompanying paper. For each training condition (e.g., depths 0–2), models are fine-tuned on the groups from those depths, then evaluated on all seven depths.

---

Versioning and maintenance

The current release is version 1.0.0. Future versions are planned to extend coverage with additional categories and more fine-grained consistency tests. Updates will be published on this repository following semantic versioning. The dataset is fully reproducible from the accompanying generation script (scripts/generate_ss_groups.py, seed 42).

---

Usage

import json

with open("ss_groups.json") as f:
    groups = json.load(f)

# Filter to a single category
capital_groups = [g for g in groups if g["category"] == "capital_word_order"]

# Inspect phrasings and expected answer
for group in capital_groups[:3]:
    print(group["expected"], group["questions"])

---

License

CC BY 4.0

---

Citation

@inproceedings{krenc2026consistency,
  title     = {Consistency Is Not Correctness: Measuring Surface-Form Sensitivity and Misgeneralization in Language Models},
  author    = {Krenc, Krzysztof},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) -- Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://huggingface.co/datasets/Hravan/semantic-sensitivity},
}