BurnyCoder's picture
Update private research reports
1f17020 verified
|
Raw History Blame Contribute Delete
1.59 kB

Dev experiments

The same legacy questions are used for each swarm size and baseline. These results select a size and estimate cost.

Phase Domain Method Graded / attempts Quality [95% CI] USD/question¹ Median / p95 s Errors
dev code arctic 10 / 10 83.1 [81.3, 85.0] 0.0032 39.1 / 136.2 0
dev code oneshot 10 / 10 78.7 [73.3, 82.9] 0.0008 7.9 / 13.3 0
dev code stock 10 / 10 81.1 [75.9, 85.3] 0.0042 78.2 / 125.3 0
dev code swarm1 10 / 10 82.6 [80.0, 85.4] 0.0041 47.6 / 111.2 0
dev code swarm2 10 / 10 71.6 [61.6, 80.6] 0.0082 98.5 / 262.4 0
dev code swarm4 10 / 10 75.1 [66.7, 82.3] 0.0172 135.3 / 238.0 0
dev code swarm8 10 / 10 78.9 [72.8, 84.2] 0.0219 112.1 / 253.9 0
dev news arctic 10 / 10 70.0% [39.7%, 89.2%] 0.0074 41.5 / 97.1 0
dev news oneshot 10 / 10 70.0% [39.7%, 89.2%] 0.0007 6.7 / 8.3 0
dev news stock 10 / 10 90.0% [59.6%, 98.2%] 0.0054 63.1 / 126.0 0
dev news swarm1 10 / 10 50.0% [23.7%, 76.3%] 0.0060 87.7 / 120.4 0
dev news swarm2 10 / 10 60.0% [31.3%, 83.2%] 0.0091 154.1 / 251.0 0
dev news swarm4 10 / 10 80.0% [49.0%, 94.3%] 0.0182 179.0 / 228.5 0
dev news swarm8 10 / 10 60.0% [31.3%, 83.2%] 0.0222 148.3 / 225.8 0

¹ Conservative inference charge using the cache-write upper rate for unclassified input tokens. Grading is included in the phase ledger.