kalyan-ks/fact-check-need-classification

Dataset Card: Fact Check Need Classification Dataset

Dataset Summary

This dataset contains curated user prompts designed to train and benchmark semantic routing classifiers that determine whether an incoming query requires factual verification (FACT_CHECK_NEEDED) or not (NO_FACT_CHECK_NEEDED).

Production LLM systems face latency and API cost bottlenecks when searching the web or querying vector databases for non-factual queries. This dataset was constructed to train lightweight semantic router models capable of fast gating decisions.

Dataset Structure

Data Instances

An example instance:

{
  "user_prompt": "Can you explain why the Treaty of Versailles contributed to the outbreak of WWII?",
  "label": "FACT_CHECK_NEEDED",
  "difficulty_tier": "Medium"
}

Data Fields

• user_prompt (string): The raw text of the user prompt.
• label (string / class_label):
  • FACT_CHECK_NEEDED: Factual query, historical/scientific inquiries, statistical questions, or claims requiring grounding.
  • NO_FACT_CHECK_NEEDED: Creative writing, code generation, tone transformations, roleplay, brainteasers, or abstract opinions.
• difficulty_tier (string): Difficulty rating (Easy, Medium, Hard) indicating ambiguity, subtlety of factual reliance, and linguistic nuances.

Data Splits & Distribution

The dataset is stratified across labels and difficulty tiers to maintain consistent distributions:

Split	Records	Ratio
Train	21070	75%
Validation	2810	10%
Test	4214	15%

• Label Balance: 50% FACT_CHECK_NEEDED / 50% NO_FACT_CHECK_NEEDED.
• Difficulty Tier: Easy (20%), Medium (40%) and Hard (40%)

---

Dataset Creation

Source Data & Generation Pipeline

1. Taxonomy & Domain Coverage: Prompts cover 120 granular topics, 3 difficulty tiers (50 nuance profiles), 19 prompt interaction formats, 4 length tiers across 11 structures, 29 user personas, 8 linguistic registers, and 9 tones.
2. Quality Filtering: Synthetic candidates underwent deterministic rule gating and multi-criteria LLM Judge scoring (label correctness, naturalness, taxonomy adherence, training signal).
3. Semantic Deduplication: Near-duplicate queries were filtered out using embedding similarity clustering.
4. Stratified Split: Splits were generated using a composite key (label_difficulty_tier).

---

How to Use

from datasets import load_dataset

# Load all splits
dataset = load_dataset("kalyan-ks/fact-check-need-classification")

print(dataset)
# Output:
# DatasetDict({
#     train: Dataset({ features: ['user_prompt', 'label', 'difficulty_tier'], num_rows: 21070 }),
#     validation: Dataset({ features: ['user_prompt', 'label', 'difficulty_tier'], num_rows: 2810 }),
#     test: Dataset({ features: ['user_prompt', 'label', 'difficulty_tier'], num_rows: 4214 })
# })

---

Considerations & Limitations

• The dataset is currently in English (en).
• Prompts are synthetic and calibrated for query-routing scenarios; real-world edge cases with adversarial jailbreaks may require additional fine-tuning.

Licensing

This dataset is distributed under the MIT License.

Citation

@misc{factcheck-need-classification-dataset-2026,
  title = {Fact Check Need Classification Dataset},
  author = {Kalyan KS},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/kalyan-ks/fact-check-need-classification}
}