datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
position-bias-training-datasets
Position-Bias Training Datasets (main 4 distributions)
Synthetic retrieval training data used to induce/control position bias in embedding
models, restricted to the four main evidence-position distributions used in the paper.
Filtered at reranker threshold t030 (signal_gap ≥ 0.30, highest label purity).
Each example is a (question, positive_doc) pair. The distributions differ only in where
the relevant evidence sits inside the positive document (beginning / middle / end),
mixed… See the full description on the dataset page: https://huggingface.co/datasets/sionic-ai/position-bias-training-datasets.ALCUNA_meta_affirmative_3_mix_position_biased_unbiasedALCUNA_meta_affirmative_3_mix_position_biased_unbiased_train