# Training record of the shipped BITLSTM weights (F2L).
# One line per pass over the first 418837176 bits the model sees in the trace.
# Model: 16 cells, 90 input columns (prev_bit included), 3585 parameters.
# lr_1e6: learning rate x 1e6 of the pass (cosine from 3000 to 30).
# delta_B: in-sample replay value, not an archive size: the cost of the
# replayed bits under the corrected minus under the uncorrected
# probability, in bytes, while the model learns during the pass
# (negative = smaller).  The uncorrected cost of the range is base_B = 18981615.2.
# rel: delta_B / base_B in percent.
pass	lr_1e6	head_B	delta_B	rel_%
1	3000	18944084.4	-37530.8	-0.197722
2	2992	18928106.4	-53508.7	-0.281898
3	2970	18924096.6	-57518.5	-0.303022
4	2932	18922242.3	-59372.9	-0.312792
5	2880	18921027.7	-60587.4	-0.319190
6	2813	18920152.4	-61462.7	-0.323801
7	2734	18919427.6	-62187.6	-0.327620
8	2642	18918842.6	-62772.6	-0.330702
9	2538	18918305.4	-63309.8	-0.333532
10	2424	18917771.7	-63843.5	-0.336344
11	2301	18917244.0	-64371.1	-0.339124
12	2169	18916703.0	-64912.2	-0.341974
13	2031	18916154.5	-65460.7	-0.344863
14	1887	18915611.1	-66004.1	-0.347726
15	1740	18915077.5	-66537.6	-0.350537
16	1590	18914532.6	-67082.6	-0.353408
17	1440	18913987.6	-67627.6	-0.356279
18	1290	18913438.1	-68177.1	-0.359174
19	1143	18912893.9	-68721.3	-0.362041
20	999	18912366.6	-69248.6	-0.364819
21	861	18911864.5	-69750.7	-0.367464
22	729	18911389.3	-70225.9	-0.369968
23	606	18910956.3	-70658.8	-0.372249
24	492	18910571.3	-71043.8	-0.374277
25	388	18910242.2	-71372.9	-0.376011
26	296	18909981.5	-71633.7	-0.377385
27	217	18909796.0	-71819.1	-0.378362
28	150	18909688.9	-71926.2	-0.378926
29	98	18909661.7	-71953.4	-0.379069
30	60	18909705.0	-71910.1	-0.378841
31	38	18909765.6	-71849.5	-0.378522
32	30	18909773.4	-71841.8	-0.378481
