SalutefaceAAC/logical-fallacy-classification

Logical Fallacy Classification

A synthetic benchmark of 183 hand-authored examples for detecting and classifying logical fallacies in natural language text. Each example is labelled with a fallacy type, a binary label (fallacious / sound), an explanation of the reasoning error, domain, difficulty level, and subcategory.

Dataset at a Glance

Field	Values
Total examples	183
Fallacy types	20 (+ none for sound examples)
Labels	fallacious, sound
Difficulty levels	subtle, moderate, overt
Domains	politics, science, health, ethics, law, economics, education, philosophy, marketing, everyday
Splits	train / validation / test
License	Apache 2.0

Fallacy Types

Fallacy Type	Count	Description
ad_hominem	12	Attacking the person, not the argument
straw_man	11	Misrepresenting someone's position
slippery_slope	10	Unjustified chain of increasingly extreme consequences
false_dilemma	10	Presenting only two options when more exist
appeal_to_authority	10	Using authority outside their domain or without evidence
hasty_generalization	10	Drawing broad conclusions from small/biased samples
post_hoc	9	Assuming causation from temporal succession
circular_reasoning	8	Using the conclusion as a premise
appeal_to_emotion	8	Manipulating emotions instead of providing evidence
appeal_to_popularity	8	What is popular must be true/right
false_equivalence	8	Treating importantly different things as the same
red_herring	8	Introducing irrelevant information to distract
appeal_to_nature	7	Natural = good; unnatural = bad
equivocation	7	Using the same word in different senses
sunk_cost	7	Continuing because of past investment, not future value
anecdotal_evidence	7	Personal experience over statistical evidence
no_true_scotsman	6	Redefining categories to exclude counterexamples
tu_quoque	6	Deflecting criticism by pointing to the accuser's faults
appeal_to_ignorance	6	Absence of disproof as proof
naturalistic_fallacy	5	Deriving 'ought' from 'is'
none (sound)	20	Non-fallacious, well-reasoned statements

Sample Records

Ad Hominem · circumstantial · overt

{
  "text": "Senator Williams says we need stricter gun laws, but she's never even held a gun. Her opinion is worthless.",
  "fallacy_type": "ad_hominem",
  "label": "fallacious",
  "explanation": "Dismisses the senator's policy argument based on her personal experience rather than engaging with the merits of gun legislation.",
  "domain": "politics",
  "difficulty": "overt",
  "subcategory": "circumstantial"
}

False Dilemma · binary_framing · overt

{
  "text": "You're either with us or you're with the terrorists.",
  "fallacy_type": "false_dilemma",
  "label": "fallacious",
  "explanation": "Presents only two mutually exclusive positions, ignoring the possibility of neutrality, nuance, or third options.",
  "domain": "politics",
  "difficulty": "overt",
  "subcategory": "binary_framing"
}

Slippery Slope · exaggerated_endpoint · subtle

{
  "text": "If we accept one typo in official government documents, accuracy will soon become completely meaningless in public life.",
  "fallacy_type": "slippery_slope",
  "label": "fallacious",
  "explanation": "Extends a trivial, isolated event to a systemic breakdown with no intermediate reasoning.",
  "domain": "everyday",
  "difficulty": "subtle",
  "subcategory": "exaggerated_endpoint"
}

Equivocation · lexical_shift · moderate

{
  "text": "All laws have exceptions. Evolution is a law of nature. Therefore evolution has exceptions.",
  "fallacy_type": "equivocation",
  "label": "fallacious",
  "explanation": "'Law' in jurisprudence (human rule) vs. scientific law (observed regularity) are different senses — the syllogism exploits this ambiguity.",
  "domain": "science",
  "difficulty": "moderate",
  "subcategory": "lexical_shift"
}

No True Scotsman · moving_goalpost · moderate

{
  "text": "True Christians don't judge others. — What about the ones who do? — They're not true Christians.",
  "fallacy_type": "no_true_scotsman",
  "label": "fallacious",
  "explanation": "Immunizes the claim against counterexample by revising the definition of 'true Christian' post-hoc.",
  "domain": "philosophy",
  "difficulty": "moderate",
  "subcategory": "moving_goalpost"
}

Appeal to Nature · naturalistic_fallacy · moderate

{
  "text": "Humans have always lived in hierarchical societies — inequality is just natural.",
  "fallacy_type": "appeal_to_nature",
  "label": "fallacious",
  "explanation": "Naturalness doesn't justify a social arrangement; many natural phenomena are harmful and we work to change them.",
  "domain": "ethics",
  "difficulty": "moderate",
  "subcategory": "naturalistic_fallacy"
}

Sound Reasoning · none · moderate

{
  "text": "Multiple peer-reviewed meta-analyses consistently show that vaccines do not cause autism. Therefore the claim that vaccines cause autism is not supported by evidence.",
  "fallacy_type": "none",
  "label": "sound",
  "explanation": "Cites convergent scientific evidence from high-quality studies; the conclusion follows from the evidence.",
  "domain": "health",
  "difficulty": "moderate",
  "subcategory": ""
}

Schema

{
  "id": int,               # unique row identifier
  "text": str,             # the statement to classify
  "fallacy_type": str,     # fallacy name (see table above) or "none"
  "label": str,            # "fallacious" | "sound"
  "explanation": str,      # why this is (or isn't) a fallacy
  "domain": str,           # subject domain
  "difficulty": str,       # "subtle" | "moderate" | "overt"
  "subcategory": str       # finer-grained fallacy variant (may be empty for "none")
}

Usage

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/logical-fallacy-classification")
train, val, test = ds["train"], ds["validation"], ds["test"]

# --- Task 1: binary detection ---
for ex in train:
    text  = ex["text"]
    label = 1 if ex["label"] == "fallacious" else 0

# --- Task 2: multi-class (20 fallacy types) ---
from collections import Counter
type_dist = Counter(ex["fallacy_type"] for ex in train)
print(type_dist.most_common(5))

# --- Difficulty-stratified split ---
subtle   = test.filter(lambda x: x["difficulty"] == "subtle")
moderate = test.filter(lambda x: x["difficulty"] == "moderate")
overt    = test.filter(lambda x: x["difficulty"] == "overt")

# --- Zero-shot LLM evaluation with anthropic SDK ---
import anthropic

client = anthropic.Anthropic()

def detect_fallacy(text: str) -> dict:
    prompt = (
        "Identify whether the following statement contains a logical fallacy. "
        "If yes, name the fallacy type and briefly explain it. "
        "If no, say 'sound reasoning'.\n\n"
        f"Statement: {text}"
    )
    msg = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=256,
        messages=[{"role": "user", "content": prompt}]
    )
    return {"response": msg.content[0].text}

# evaluate on the test split
results = [detect_fallacy(ex["text"]) for ex in test]

Splits

Split	Size
train	137
validation	23
test	23

Intended Use

• Fine-tuning and benchmarking classifiers for logical fallacy detection
• Evaluating large language model reasoning and critical thinking
• Educational tools for rhetoric and argumentation courses
• Fact-checking and misinformation detection pipeline components

Methodology

All 183 examples were hand-authored by a domain expert with coverage of:

• 20 canonical logical fallacy types from informal logic
• 10 subject domains (politics, science, health, law, economics, ethics, philosophy, education, marketing, everyday)
• 3 difficulty levels: overt (easily spotted), moderate (requires attention), subtle (requires domain knowledge)
• 20 sound (non-fallacious) examples to prevent trivial detection biases
• Subcategory labels for finer-grained analysis within each fallacy type

Limitations

• Scale: 183 examples is a benchmark, not a training corpus. Use for evaluation, few-shot prompting, or augmentation seeding — not standalone fine-tuning.
• Synthetic construction: All examples are author-controlled and "clean"; real-world fallacies are typically embedded in longer discourse with shared context.
• Class imbalance: 89% fallacious / 11% sound. Report balanced accuracy (macro-averaged recall) alongside raw accuracy to avoid trivial "always-fallacious" baselines.
• Boundary cases excluded: Sound examples are clearly sound; fallacious examples are clearly fallacious. Arguments that are fallacious only in context are not included.
• Sound class too small (n=20) for rigorous cross-domain analysis of non-fallacious reasoning.
• English only: Fallacy recognition is culturally situated; some types manifest differently across languages and rhetorical traditions.
• Subcategory incompleteness: Subcategories cover common variants but are not exhaustive within each fallacy type.

Evaluation Guidance

For binary detection, always report balanced accuracy (macro-averaged recall) to avoid inflated scores from the class imbalance. A trivial "always-fallacious" classifier achieves 91.3% accuracy but only 50% balanced accuracy.

For multi-class identification, use macro-F1 over the 20 types. Types with ≤6 examples (no_true_scotsman, tu_quoque, appeal_to_ignorance, naturalistic_fallacy) will show high variance — report confidence intervals or aggregate them in significance tests.

For difficulty-stratified analysis, report separately for overt / moderate / subtle subsets to reveal where a model fails rather than just overall accuracy.

See REPORT.md for full evaluation protocols and comparison with related datasets.

Related Work

• LOGIC (Jin et al., 2022) — 13,000 natural Reddit instances across 13 fallacy types; higher ecological validity, noisier labels. Complementary to this dataset.
• MAFALDA (Helwe et al., 2023) — French-language fallacy benchmark (200+ examples, 15 types).
• ArgMining shared tasks (2014–2023) — upstream argument quality and stance tasks at major NLP venues.
• FallacyDet (Goffredo et al., 2022) — political debate fallacy detection with pre-trained transformers.

Citation

@dataset{salutefaceaac2026fallacies,
  title     = {Logical Fallacy Classification},
  author    = {SalutefaceAAC},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/SalutefaceAAC/logical-fallacy-classification},
  license   = {Apache 2.0}
}