Preference pairs for DPO fine-tuning of Qwen/Qwen2.5-0.5B-Instruct against the Oculus multilingual AI text detector on Spanish academic abstracts. Repository id: pymlex/ai-generated-texts.

	Dataset size

Statistic Count

Train abstracts processed 8891

DPO pairs retained 6396

Pairs skipped by logit margin 2495

Empty paraphrase pairs 0

Logit margin threshold: absolute gap at least 1.… See the full description on the dataset page: https://huggingface.co/datasets/pymlex/ai-generated-texts.