Halluguard Preferences 76K

HalluGuard-Prefs is a 76,708-entry synthetic preference dataset created by author lrsbrgrn for training models to detect hallucinations in text. It was used to fine-tune the HalluGuard-Qwen3-4B model via Odds Ratio Preference Optimization (ORPO). The dataset was introduced in a paper at the 64th Annual Meeting of the Association and was last updated on Hugging Face in June 2026.

Use Cases

• Fine-tuning LLMs for hallucination detection based on the described preference tuples.
• Training models to reason and justify hallucination judgments based on the dataset's design.
• Evaluating model performance on document-grounded factual consistency tasks.
• Researching Odds Ratio Preference Optimization (ORPO) and other alignment techniques.

Strengths

• Contains 76,708 high-quality preference tuples.
• Designed specifically for teaching models to reason about hallucination detection.
• Used to successfully fine-tune a 4B-parameter model (HalluGuard-Qwen3-4B).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is synthetic, which may limit its generalizability to real-world text distributions.
• Row count is known, but other structural details like file formats and sample data are unavailable.

Provenance

Source

Hugging Face, uploaded by lrsbrgrn.

Collection Method

Synthetically generated for model training.

Time Range

null

Freshness

Last updated 2026-06-24 08:22:30.

Geography

null