SPARK training data for context–memory conflicts
SPARK releases two sets of training and validation records for a policy that regulates how much an LLM relies on parametric knowledge during retrieval-augmented generation. The two archives correspond to Qwen3-1.7B and Qwen3-8B.
Which examples are in each archive?
Both ZIPs contain train.json and valid.json. The Qwen3-1.7B release has 9,407 training and 2,351 validation records. The Qwen3-8B release has 14,265 and 3,565. Each example includes a question, context, a counterfactual context, the model’s parametric answer and a context answer. The field is_parametric_answer_right records whether the parametric answer is correct.
The authors used FlashRAG and MRQA to synthesize the training data. Use the archive corresponding to the model whose parametric answers you want to study. Data preparation in the authors’ repository
Data files
Authors’ Hugging Face dataset card · The downloads below use a fixed revision.
Related paper
SPARK: Adaptive Parametric Knowledge Participation for Faithful Retrieval-Augmented Generation · Paper on arXiv