1{2 "document_id": "lb_qasper_677dad703988e8f26b322211803fad38ffe9e58d1725f51a",3 "family": "public_qasper",4 "split": "test",5 "document_hash": "5724368043143d51f10254b9efc23d824840c3605c32f9c85bacfa4c4ca5faa4",6 "questions": [7 {8 "question_id": "q00",9 "qtype": "qasper",10 "question": "How are weights dynamically adjusted?",11 "answer": "One can think $(1-p_{i1})$ as a weight associated with each example, which changes as training proceeds. The intuition of changing $p_{i1}$ to $(1-p_{i1}) p_{i1}$ is to push down the weight of easy examples. For easy examples whose probability are approaching 0 or 1, $(1-p_{i1}) p_{i1}$ makes the model attach significantly less focus to them. Figure FIGREF23 gives gives an explanation from the perspective in derivative: the derivative of $\\frac{(1-p)p}{1+(1-p)p}$ with respect to $p$ approaches 0 immediately after $p$ approaches 0, which means the model attends less to examples once they are correctly classified.",12 "aliases": [13 "associates each training example with a weight in proportion to $(1-p)$, and this weight dynamically changes as training proceeds"14 ],15 "answer_type": "qa_f1",16 "tolerance": 0.0,17 "supporting_fact_ids": [],18 "answer_position_frac": -1.019 }20 ],21 "facts": []22}