Datasetpublic

stindardlogic/hallucination-grounding-dpo-4k

Hallucination Grounding DPO Pairs (4K) DPO preference pairs targeting the full spectrum of factuality failures — from hallucination to over-hedging. Motivation Existing refusal/safety datasets focus on what not to say. This dataset targets the orthogonal challenge: when to say "I don't know" vs. when to answer confidently. Models that over-refuse waste user trust; models that hallucinate destroy it. Dataset Description 4,000 preference pairs across… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/hallucination-grounding-dpo-4k.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes33downloads

Dataset Card

Hallucination Grounding DPO Pairs (4K)

DPO preference pairs targeting the full spectrum of factuality failures — from hallucination to over-hedging.

Motivation

Existing refusal/safety datasets focus on what not to say. This dataset targets the orthogonal challenge: when to say "I don't know" vs. when to answer confidently. Models that over-refuse waste user trust; models that hallucinate destroy it.

Dataset Description

4,000 preference pairs across 3 scenario types:

Type	Count	Description
A — Should admit uncertainty	~2,000	Rejected = confident hallucination of unknowable fact
B — Should answer confidently	~1,000	Rejected = unnecessary hedging on well-known fact
C — Should give partial answer	~1,000	Rejected = fabricated false precision

Rejection failure modes:

—hallucination — confidently states specific false facts (dates, quotes, numbers)
—over_hedging — adds excessive uncertainty to well-established knowledge
—false_precision — provides invented specific numbers/details instead of approximate ranges

Format

json

{
  "prompt": "What did Elon Musk tweet on March 15, 2023 at 9:47 AM?",
  "chosen": "I don't have access to specific tweets by timestamp...",
  "rejected": "On March 15, 2023 at 9:47 AM, Elon Musk tweeted: 'The bird is freed...'",
  "metadata": {"type": "A", "rejection_type": "hallucination"},
  "id": "abc123"
}

Use Case

—Fine-tune models to calibrate confidence appropriately
—Reduce both hallucination AND over-refusal simultaneously
—Train models on epistemic humility (knowing what they don't know)
—Complement `stindardlogic/refusal-overrefusal-50k`

Design Principle

The chosen response never hallucates, never claims unknowable precision, but also never refuses to answer well-established facts. The rejected response always commits one of the three failure modes.

License

Apache 2.0