{ "actions": [ "A1", "A2|budget_tokens=512", "A3|budget_tokens=512|blind", "A3|budget_tokens=512|aware", "A4|budget_tokens=512|blind", "A5|budget_tokens=512|blind", "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "A7|codec=zstd", "A8|budget_tokens=512|blind", "B0", "B1", "B2|bits=8", "B3|keep_frac=0.25", "B5|keep_frac=0.25|aware", "B9|bits=8|keep_frac=0.25|policy=attention|aware" ], "cost_col": "cost_amortized_ms", "tau": 0.9, "table": { "exec|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 0.5, "q_lo": 0.0, "cost": 1329.8871670267545, "n": 2 }, { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1476.8745410256088, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1518.9385620178655, "n": 2 } ] }, "exec|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "A3|budget_tokens=512|blind", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 1631.058249855414, "n": 2 }, { "action": "A2|budget_tokens=512", "q_mean": 1.0, "q_lo": 1.0, "cost": 1417.3036039574072, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 1.0, "q_lo": 1.0, "cost": 1332.8896454768255, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.5, "q_lo": 0.0, "cost": 8919.545584125444, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 8856.286415655632, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 1261.0194170265459, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 1632.7006250503473, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 2142.073229479138, "n": 2 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 1853.2067709602416, "n": 2 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 1809.8062080680393, "n": 2 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 1990.0899369386025, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.0, "q_lo": 0.0, "cost": 1653.049167478457, "n": 2 } ] }, "exec|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 1885.1518545416184, "n": 2 }, { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.5, "q_lo": 0.0, "cost": 1798.074854479637, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1998.245208291337, "n": 2 } ] }, "exec|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "A3|budget_tokens=512|blind", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 2106.2395630870014, "n": 2 }, { "action": "A2|budget_tokens=512", "q_mean": 1.0, "q_lo": 1.0, "cost": 1777.6976449531503, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 1.0, "q_lo": 1.0, "cost": 1548.3912916388363, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.5, "q_lo": 0.0, "cost": 7460.223854461219, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 6980.376125022303, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 1045.0055415276438, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 1939.9121680180542, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 6249.4903958868235, "n": 2 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 2214.8592923767865, "n": 2 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 2213.8642499921843, "n": 2 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 2394.7104800026864, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.0, "q_lo": 0.0, "cost": 1192.3931669443846, "n": 2 } ] }, "public_2wikimqa|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "B5|keep_frac=0.25|aware", "best_static": "B5|keep_frac=0.25|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1401.3548743678257, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1821.9773949240334, "n": 2 } ] }, "public_2wikimqa|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "A1", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 0.0, "q_lo": 0.0, "cost": 958.6374779464677, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 0.0, "q_lo": 0.0, "cost": 1236.2463335739449, "n": 2 }, { "action": "B0", "q_mean": 0.0, "q_lo": 0.0, "cost": 1379.270581412129, "n": 2 }, { "action": "B1", "q_mean": 0.0, "q_lo": 0.0, "cost": 1497.370915429201, "n": 2 }, { "action": "B2|bits=8", "q_mean": 0.0, "q_lo": 0.0, "cost": 1899.9604169512168, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.5, "q_lo": 0.0, "cost": 1568.692895467393, "n": 2 } ] }, "public_2wikimqa|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "B5|keep_frac=0.25|aware", "best_static": "B5|keep_frac=0.25|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1259.7939987899736, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 1319.705936824903, "n": 2 } ] }, "public_2wikimqa|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "B3|keep_frac=0.25", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 0.0, "q_lo": 0.0, "cost": 1315.37816679338, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 0.0, "q_lo": 0.0, "cost": 1215.5741045135073, "n": 2 }, { "action": "B0", "q_mean": 0.0, "q_lo": 0.0, "cost": 1290.58462503599, "n": 2 }, { "action": "B1", "q_mean": 0.0, "q_lo": 0.0, "cost": 1291.3224364165217, "n": 2 }, { "action": "B2|bits=8", "q_mean": 0.0, "q_lo": 0.0, "cost": 1230.071353900712, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.0, "q_lo": 0.0, "cost": 1153.9108945289627, "n": 2 } ] }, "public_2wikimqa|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 140.73666697368026, "n": 2 } ] }, "public_2wikimqa|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "best_static": "A2|budget_tokens=512", "candidates": [ { "action": "A2|budget_tokens=512", "q_mean": 0.0, "q_lo": 0.0, "cost": 136.72804203815758, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 112.30885348049924, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 9116.805187019054, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 9143.03629292408, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 101.88310500234365, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 728.3446455257945, "n": 2 } ] }, "public_2wikimqa|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 0.0, "q_lo": 0.0, "cost": 178.05522889830172, "n": 2 } ] }, "public_2wikimqa|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "best_static": "A2|budget_tokens=512", "candidates": [ { "action": "A2|budget_tokens=512", "q_mean": 0.0, "q_lo": 0.0, "cost": 116.48358352249488, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.5, "q_lo": 0.0, "cost": 131.05783401988447, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 8600.43547960231, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 8467.526854481548, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 112.01987491222098, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 1496.8782920041122, "n": 2 } ] }, "public_qasper|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "B5|keep_frac=0.25|aware", "best_static": "B5|keep_frac=0.25|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 722.1069170045666, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 780.5135424714535, "n": 2 } ] }, "public_qasper|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "A1", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 595.0978129403666, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 604.6264575561509, "n": 2 }, { "action": "B0", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 1024.1511670174077, "n": 2 }, { "action": "B1", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 991.0490409820341, "n": 2 }, { "action": "B2|bits=8", "q_mean": 0.24999999999999997, "q_lo": 0.125, "cost": 765.4312709346414, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.0625, "q_lo": 0.0, "cost": 600.5491039832123, "n": 2 } ] }, "public_qasper|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "best_static": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.2, "q_lo": 0.0, "cost": 711.4740828401409, "n": 2 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.2625, "q_lo": 0.125, "cost": 781.3643955742009, "n": 2 } ] }, "public_qasper|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "B3|keep_frac=0.25", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 0.18749999999999997, "q_lo": 0.0, "cost": 722.5117080379277, "n": 2 }, { "action": "A7|codec=zstd", "q_mean": 0.18749999999999997, "q_lo": 0.0, "cost": 720.7211870700121, "n": 2 }, { "action": "B0", "q_mean": 0.18749999999999997, "q_lo": 0.0, "cost": 1075.719332555309, "n": 2 }, { "action": "B1", "q_mean": 0.18749999999999997, "q_lo": 0.0, "cost": 866.128791065421, "n": 2 }, { "action": "B2|bits=8", "q_mean": 0.0625, "q_lo": 0.0, "cost": 609.4338131370023, "n": 2 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.0, "q_lo": 0.0, "cost": 524.0459991036914, "n": 2 } ] }, "public_qasper|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 0.2932692307692307, "q_lo": 0.125, "cost": 209.8669374245219, "n": 2 } ] }, "public_qasper|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "A4|budget_tokens=512|blind", "best_static": "A4|budget_tokens=512|blind", "candidates": [ { "action": "A2|budget_tokens=512", "q_mean": 0.0625, "q_lo": 0.0, "cost": 156.89179295441136, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 136.68591663008556, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.2932692307692307, "q_lo": 0.125, "cost": 8117.365646059625, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 8089.817729953211, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 118.74318757327273, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 519.0687498543411, "n": 2 } ] }, "public_qasper|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 0.19763513513513511, "q_lo": 0.125, "cost": 298.25204209191725, "n": 2 } ] }, "public_qasper|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "best_static": "A2|budget_tokens=512", "candidates": [ { "action": "A2|budget_tokens=512", "q_mean": 0.0, "q_lo": 0.0, "cost": 163.6342714773491, "n": 2 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 131.36837491765618, "n": 2 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 3503.9038544055074, "n": 2 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 6761.200083012227, "n": 2 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 117.29304090840742, "n": 2 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.15789473684210525, "q_lo": 0.0, "cost": 934.2439590254799, "n": 2 } ] }, "synthetic|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "B5|keep_frac=0.25|aware", "best_static": "B5|keep_frac=0.25|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 2380.357041885145, "n": 5 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 2720.084957452491, "n": 5 } ] }, "synthetic|medium|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "B2|bits=8", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 2029.3159250169992, "n": 5 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 1883.6203998420388, "n": 5 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 901.3801999855787, "n": 5 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 951.6737081110477, "n": 5 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 798.1130251660943, "n": 5 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.2, "q_lo": 0.2, "cost": 692.6723245764151, "n": 5 } ] }, "synthetic|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "B5|keep_frac=0.25|aware", "best_static": "B5|keep_frac=0.25|aware", "candidates": [ { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.8, "q_lo": 0.8, "cost": 2617.682141601108, "n": 5 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.8, "q_lo": 0.8, "cost": 2645.6742170033976, "n": 5 } ] }, "synthetic|medium|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "B0", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 2015.7257752027363, "n": 5 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 3648.054808191955, "n": 5 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 926.8670671852306, "n": 5 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 964.7658923873678, "n": 5 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 955.3255414124578, "n": 5 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.2, "q_lo": 0.2, "cost": 789.7272833390161, "n": 5 } ] }, "synthetic|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 428.00330000463873, "n": 10 }, { "action": "B5|keep_frac=0.25|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 1493.95510929171, "n": 5 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 1567.7744757616892, "n": 5 } ] }, "synthetic|short|Llama-3.2-3B-Instruct-bf16->Llama-3.2-3B-Instruct-bf16|blind": { "chosen": "B1", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 774.6145416051149, "n": 5 }, { "action": "A2|budget_tokens=512", "q_mean": 0.3, "q_lo": 0.2, "cost": 341.9635915895924, "n": 10 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.4, "q_lo": 0.4, "cost": 363.9699418679811, "n": 10 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.3, "q_lo": 0.2, "cost": 4784.523370966781, "n": 10 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.0, "q_lo": 0.0, "cost": 4584.249533410184, "n": 10 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.3, "q_lo": 0.2, "cost": 2474.1135540185496, "n": 10 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 1710.8100838027894, "n": 5 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.2, "q_lo": 0.0, "cost": 2308.3051127032377, "n": 10 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 714.2056331736967, "n": 5 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 599.7976336162537, "n": 5 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 622.288533882238, "n": 5 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.2, "q_lo": 0.2, "cost": 638.1709832465276, "n": 5 } ] }, "synthetic|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|aware": { "chosen": "A3|budget_tokens=512|aware", "best_static": "A3|budget_tokens=512|aware", "candidates": [ { "action": "A3|budget_tokens=512|aware", "q_mean": 1.0, "q_lo": 1.0, "cost": 492.4797534924717, "n": 13 }, { "action": "B5|keep_frac=0.25|aware", "q_mean": 0.8, "q_lo": 0.8, "cost": 1461.255682632327, "n": 5 }, { "action": "B9|bits=8|keep_frac=0.25|policy=attention|aware", "q_mean": 0.8, "q_lo": 0.8, "cost": 1528.5088498611003, "n": 5 } ] }, "synthetic|short|Qwen2.5-3B-Instruct-bf16->Qwen2.5-3B-Instruct-bf16|blind": { "chosen": "B1", "best_static": "A1", "candidates": [ { "action": "A1", "q_mean": 1.0, "q_lo": 1.0, "cost": 834.2573178815655, "n": 8 }, { "action": "A2|budget_tokens=512", "q_mean": 0.38461538461538464, "q_lo": 0.2, "cost": 335.3575096788028, "n": 13 }, { "action": "A3|budget_tokens=512|blind", "q_mean": 0.46153846153846156, "q_lo": 0.4, "cost": 379.59068921358823, "n": 13 }, { "action": "A4|budget_tokens=512|blind", "q_mean": 0.5384615384615384, "q_lo": 0.4, "cost": 3687.5417852648893, "n": 13 }, { "action": "A5|budget_tokens=512|blind", "q_mean": 0.23076923076923078, "q_lo": 0.2, "cost": 4437.3677050999295, "n": 13 }, { "action": "A6|budget_tokens=512|max_fetch_bytes=2000|max_lookups=6|blind", "q_mean": 0.23076923076923078, "q_lo": 0.2, "cost": 5032.344285452452, "n": 13 }, { "action": "A7|codec=zstd", "q_mean": 1.0, "q_lo": 1.0, "cost": 2361.7878749937518, "n": 8 }, { "action": "A8|budget_tokens=512|blind", "q_mean": 0.2, "q_lo": 0.0, "cost": 29813.514716585632, "n": 10 }, { "action": "B0", "q_mean": 1.0, "q_lo": 1.0, "cost": 798.3306667534634, "n": 5 }, { "action": "B1", "q_mean": 1.0, "q_lo": 1.0, "cost": 742.7887752186507, "n": 5 }, { "action": "B2|bits=8", "q_mean": 1.0, "q_lo": 1.0, "cost": 804.3182000052184, "n": 5 }, { "action": "B3|keep_frac=0.25", "q_mean": 0.2, "q_lo": 0.2, "cost": 510.90556636918336, "n": 5 } ] } }, "global_best_static": "A1", "fitted_on": [ "E1_pilot.jsonl" ], "n_dev_rows": 1037, "global_best_static_by_access": { "aware": "A3|budget_tokens=512|aware", "blind": "A1" } }