table4_e3_qa_refusal.csv

Model,RAG Context,Answered Rate (%),Citation P (%),Citation R (%),Citation F1 (%),Refusal P / R / F1 (%),Exact Refusal Format (%),Aux Token-F1 (Dev Only),TTFT p50 / p95 (ms),Latency p50 / p95 (ms),Tok/s (mean),Peak VRAM (MB),Failure Rate (%)
Qwen/Qwen3.5-4B (Report Baseline),no_rag,0.7,"0.0 [0.0, 0.0]","0.0 [0.0, 0.0]","0.0 [0.0, 0.0]",33.5 / 100.0 / 50.2,100.0,0.2,200 / 236,482 / 634,30.7,4013.0,0.0
Qwen/Qwen3.5-4B (Report Baseline),best_dense,100.0,"100.0 [100.0, 100.0]","83.0 [80.3, 85.7]","88.7 [86.9, 90.4]",100.0 / 84.7 / 91.7,99.2,48.6,724 / 1121,1875 / 3809,29.2,4013.0,0.0
Qwen/Qwen3.5-4B (Report Baseline),best_hybrid,100.0,"100.0 [100.0, 100.0]","84.7 [82.0, 87.3]","89.8 [88.0, 91.6]",100.0 / 85.3 / 92.1,98.4,48.1,721 / 1108,1936 / 3538,29.4,4013.0,0.0
Qwen/Qwen3.5-4B (Report Baseline),best_hybrid_reranker,100.0,"100.0 [100.0, 100.0]","86.3 [83.7, 88.8]","90.9 [89.1, 92.6]",100.0 / 84.7 / 91.7,99.2,47.6,559 / 994,1710 / 3677,31.2,4013.0,0.0
Qwen/Qwen3-4B-Instruct-2507 (Supplementary),no_rag,0.3,"0.0 [0.0, 0.0]","0.0 [0.0, 0.0]","0.0 [0.0, 0.0]",33.4 / 100.0 / 50.1,99.6,0.1,48 / 56,294 / 303,54.4,3845.0,0.0
Qwen/Qwen3-4B-Instruct-2507 (Supplementary),best_dense,96.7,"81.7 [77.5, 85.7]","71.3 [67.3, 75.2]","74.2 [70.3, 77.9]",93.4 / 94.0 / 93.7,100.0,48.7,314 / 865,1135 / 3170,39.3,3845.0,0.0
Qwen/Qwen3-4B-Instruct-2507 (Supplementary),best_hybrid,100.0,"86.7 [83.2, 90.2]","74.7 [70.7, 78.3]","78.1 [74.7, 81.3]",100.0 / 94.0 / 96.9,100.0,50.3,304 / 780,1308 / 2855,40.7,3845.0,0.0
Qwen/Qwen3-4B-Instruct-2507 (Supplementary),best_hybrid_reranker,100.0,"93.0 [90.0, 95.7]","77.7 [74.0, 81.2]","82.8 [79.4, 85.7]",100.0 / 94.0 / 96.9,100.0,49.9,234 / 524,1126 / 2777,43.0,3845.0,0.0