robbiethompson2018 commited on
Commit
aa5dfba
·
verified ·
1 Parent(s): 51f657d

Add 507-fable_ft-token-batches_s1

Browse files
runs/507-fable_ft-token-batches_s1/eval_dev.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
runs/507-fable_ft-token-batches_s1/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8812a140b2f4eecea299ab466cd38b8b02f6ec531de9a3820f655ed32018f73f
3
+ size 25624925
runs/507-fable_ft-token-batches_s1/found_targets.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
runs/507-fable_ft-token-batches_s1/lean_gate.disagree.jsonl ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ {"prompt": "THM ( P v R ) SEQ ( P > ( R > ( ( ~ ( ~ R ) ) > ( P & ( ~ ( ~ R ) ) ) ) ) ) PRF", "lean_text": "have n39 : ( P → ( R → ( ( ¬ ( ¬ R ) ) → ( P ∧ ( ¬ ( ¬ R ) ) ) ) ) ) := ( fun ( n40 : P ) => by have n41 : ( R → ( ( ¬ ( ¬ R ) ) → ( P ∧ ( ¬ ( ¬ R ) ) ) ) ) := ( fun ( n42 : R ) => by have n43 : ( ( ¬ ( ¬ R ) ) → ( P ∧ ( ¬ ( ¬ R ) ) ) ) := ( fun ( n44 : ( ¬ ( ¬ R ) ) ) => by have n45 : ( P ∧ ( ¬ ( ¬ R ) ) ) := ⟨ n40 , n44 ⟩ ; exact n45 ) ; exact n43 ) ; exact n41 ) ; exact n39", "nd": "N1 | P : AS ; N2 | | R : AS ; N3 | | | ( ~ ( ~ R ) ) : AS ; N4 | | | ( P & ( ~ ( ~ R ) ) ) : ANDI N1 N3 ; N5 | | ( ( ~ ( ~ R ) ) > ( P & ( ~ ( ~ R ) ) ) ) : IMPI N3 N4 ; N6 | ( R > ( ( ~ ( ~ R ) ) > ( P & ( ~ ( ~ R ) ) ) ) ) : IMPI N2 N5 ; N7 ( P > ( R > ( ( ~ ( ~ R ) ) > ( P & ( ~ ( ~ R ) ) ) ) ) ) : IMPI N1 N6 ; QED", "nd_ok": false, "lean_ok": true}
2
+ {"prompt": "THM Q , R , ( Q v Q ) SEQ ( ( ~ Q ) > ( ( ~ ( ~ Q ) ) > ( Q > ( ( R v ( ~ R ) ) & ( ~ ( ~ Q ) ) ) ) ) ) PRF", "lean_text": "have n21 : Q := h1 ; have n22 : R := h2 ; have n23 : ( Q ∨ Q ) := Or.inr n21 ; have n24 : ( ( ¬ Q ) → ( ( ¬ ( ¬ Q ) ) → ( Q → ( ( R ∨ ( ¬ R ) ) ∧ ( ¬ ( ¬ Q ) ) ) ) ) ) := ( fun ( n25 : ( ¬ Q ) ) => by have n26 : ( ( ¬ ( ¬ Q ) ) → ( Q → ( ( R ∨ ( ¬ R ) ) ∧ ( ¬ ( ¬ Q ) ) ) ) ) := ( fun ( n27 : ( ¬ ( ¬ Q ) ) ) => by have n28 : ( Q → ( ( R ∨ ( ¬ R ) ) ∧ ( ¬ ( ¬ Q ) ) ) ) := ( fun ( n29 : Q ) => by have n30 : ( R ∨ ( ¬ R ) ) := Or.inl n22 ; have n31 : ( ( R ∨ ( ¬ R ) ) ∧ ( ¬ ( ¬ Q ) ) ) := ⟨ n30 , n27 ⟩ ; exact n31 ) ; exact n28 ) ; exact n26 ) ; exact n24", "nd": "N1 Q : PR ; N2 R : PR ; N3 ( Q v Q ) : ORI2 N1 ; N4 | ( ~ Q ) : AS ; N5 | | ( ~ ( ~ Q ) ) : AS ; N6 | | | Q : AS ; N7 | | | ( R v ( ~ R ) ) : ORI1 N2 ; N8 | | | ( ( R v ( ~ R ) ) & ( ~ ( ~ Q ) ) ) : ANDI N7 N5 ; N9 | | ( Q > ( ( R v ( ~ R ) ) & ( ~ ( ~ Q ) ) ) ) : IMPI N6 N8 ; N10 | ( ( ~ ( ~ Q ) ) > ( Q > ( ( R v ( ~ R ) ) & ( ~ ( ~ Q ) ) ) ) ) : IMPI N5 N9 ; N11 ( ( ~ Q ) > ( ( ~ ( ~ Q ) ) > ( Q > ( ( R v ( ~ R ) ) & ( ~ ( ~ Q ) ) ) ) ) ) : IMPI N4 N10 ; QED", "nd_ok": false, "lean_ok": true}
runs/507-fable_ft-token-batches_s1/lean_gate.jsonl ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"utc": "2026-10-07T20:33:45Z", "samples": 2000, "parse_fail": 169, "distinct_checked": 1831, "both_ok": 1772, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 59, "parse_reasons": {"expected )": 128, "expected :=": 16, "formula": 12, "unbound": 7, "no-eos": 4, "name": 2}, "lean_wall_s": 3.102205514907837, "lean_proc_s": 13.859758377075195, "nd_verify_s": 0.10286211967468262, "workers": 96, "chunk": 400}
2
+ {"utc": "2026-10-07T20:34:28Z", "samples": 12000, "parse_fail": 8989, "distinct_checked": 2929, "both_ok": 76, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2853, "parse_reasons": {"unbound": 3930, "expected )": 2099, "formula": 1194, "expected :=": 1088, "no-eos": 246, "expected (": 189, "exact not last": 112, "name": 108, "PR position": 12, "expected False": 4, "expected =>": 3, "expected ;": 2, "expected None": 1, "expected fun": 1}, "lean_wall_s": 9.557167291641235, "lean_proc_s": 69.56554818153381, "nd_verify_s": 0.19178533554077148, "workers": 96, "chunk": 400}
3
+ {"utc": "2026-10-07T20:35:24Z", "samples": 11736, "parse_fail": 8310, "distinct_checked": 3103, "both_ok": 156, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2947, "parse_reasons": {"unbound": 4389, "expected )": 1917, "expected :=": 934, "formula": 395, "expected (": 248, "name": 175, "no-eos": 156, "exact not last": 68, "expected =>": 11, "expected False": 7, "PR position": 5, "expected None": 2, "expected ;": 2, "expected fun": 1}, "lean_wall_s": 9.990922927856445, "lean_proc_s": 77.43920278549194, "nd_verify_s": 0.13767504692077637, "workers": 96, "chunk": 400}
4
+ {"utc": "2026-10-07T20:36:45Z", "samples": 11736, "parse_fail": 7750, "distinct_checked": 3634, "both_ok": 425, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3209, "parse_reasons": {"unbound": 4673, "expected )": 1707, "expected :=": 578, "formula": 292, "name": 171, "expected (": 157, "no-eos": 105, "exact not last": 50, "PR position": 8, "expected =>": 3, "expected fun": 2, "expected False": 1, "expected ;": 1, "ORE box": 1, "expected None": 1}, "lean_wall_s": 11.129019260406494, "lean_proc_s": 99.38869333267212, "nd_verify_s": 0.16463446617126465, "workers": 96, "chunk": 400}
5
+ {"utc": "2026-10-07T20:38:32Z", "samples": 11712, "parse_fail": 7769, "distinct_checked": 3619, "both_ok": 421, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3198, "parse_reasons": {"unbound": 4472, "expected )": 1832, "expected :=": 749, "formula": 247, "name": 198, "expected (": 115, "exact not last": 63, "no-eos": 57, "expected ;": 9, "PR position": 7, "expected None": 6, "expected fun": 6, "expected =>": 3, "expected False": 2, "expected :": 2, "ORE box": 1}, "lean_wall_s": 12.186774969100952, "lean_proc_s": 109.03744411468506, "nd_verify_s": 0.1767113208770752, "workers": 96, "chunk": 400}
6
+ {"utc": "2026-10-07T20:40:12Z", "samples": 12000, "parse_fail": 2190, "distinct_checked": 9471, "both_ok": 3642, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 5828, "parse_reasons": {"unbound": 1543, "formula": 218, "expected )": 206, "expected (": 69, "exact not last": 47, "no-eos": 40, "name": 39, "expected ;": 18, "ORE box": 5, "PR position": 3, "expected :=": 2}, "lean_wall_s": 26.13907480239868, "lean_proc_s": 613.2161061763763, "nd_verify_s": 0.4941694736480713, "workers": 96, "chunk": 400}
7
+ {"utc": "2026-10-07T20:40:46Z", "samples": 6672, "parse_fail": 1462, "distinct_checked": 4709, "both_ok": 277, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4432, "parse_reasons": {"unbound": 1054, "name": 128, "expected )": 108, "expected (": 63, "formula": 62, "expected ;": 20, "PR position": 11, "exact not last": 9, "no-eos": 6, "expected :=": 1}, "lean_wall_s": 17.18265986442566, "lean_proc_s": 202.47911953926086, "nd_verify_s": 0.2471027374267578, "workers": 96, "chunk": 400}
8
+ {"utc": "2026-10-07T20:41:29Z", "samples": 6672, "parse_fail": 1219, "distinct_checked": 4951, "both_ok": 548, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4403, "parse_reasons": {"unbound": 927, "formula": 58, "expected )": 54, "name": 49, "expected (": 45, "PR position": 32, "no-eos": 19, "exact not last": 17, "expected ;": 15, "expected :=": 3}, "lean_wall_s": 18.542553186416626, "lean_proc_s": 228.7125232219696, "nd_verify_s": 0.24302220344543457, "workers": 96, "chunk": 400}
9
+ {"utc": "2026-10-07T20:42:29Z", "samples": 6648, "parse_fail": 1405, "distinct_checked": 4767, "both_ok": 571, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4196, "parse_reasons": {"unbound": 997, "name": 135, "formula": 107, "expected (": 54, "expected )": 46, "exact not last": 20, "no-eos": 16, "PR position": 9, "expected ;": 9, "expected :=": 8, "ORE box": 4}, "lean_wall_s": 16.68186664581299, "lean_proc_s": 197.73526620864868, "nd_verify_s": 0.25096797943115234, "workers": 96, "chunk": 400}
10
+ {"utc": "2026-10-07T20:44:23Z", "samples": 12000, "parse_fail": 1320, "distinct_checked": 10292, "both_ok": 5306, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4986, "parse_reasons": {"unbound": 931, "formula": 138, "expected (": 52, "no-eos": 45, "exact not last": 35, "expected )": 34, "expected :=": 23, "name": 20, "expected ;": 19, "ORE box": 14, "trailing tokens": 4, "PR position": 3, "expected =>": 2}, "lean_wall_s": 28.021997213363647, "lean_proc_s": 712.082484960556, "nd_verify_s": 0.5603964328765869, "workers": 96, "chunk": 400}
11
+ {"utc": "2026-10-07T20:44:53Z", "samples": 4872, "parse_fail": 915, "distinct_checked": 3627, "both_ok": 288, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3339, "parse_reasons": {"unbound": 591, "expected (": 99, "name": 64, "formula": 55, "no-eos": 34, "expected )": 29, "expected :=": 16, "expected ;": 12, "exact not last": 11, "ORE box": 3, "PR position": 1}, "lean_wall_s": 13.61846375465393, "lean_proc_s": 122.73236989974976, "nd_verify_s": 0.17999505996704102, "workers": 96, "chunk": 400}
12
+ {"utc": "2026-10-07T20:45:28Z", "samples": 4872, "parse_fail": 881, "distinct_checked": 3645, "both_ok": 397, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3248, "parse_reasons": {"unbound": 664, "formula": 65, "expected (": 64, "name": 62, "expected )": 16, "ORE box": 3, "exact not last": 2, "expected ;": 2, "expected :=": 1, "PR position": 1, "no-eos": 1}, "lean_wall_s": 13.466618061065674, "lean_proc_s": 119.35129308700562, "nd_verify_s": 0.2013263702392578, "workers": 96, "chunk": 400}
13
+ {"utc": "2026-10-07T20:46:22Z", "samples": 4872, "parse_fail": 1096, "distinct_checked": 3504, "both_ok": 492, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3012, "parse_reasons": {"unbound": 770, "name": 87, "formula": 82, "expected (": 54, "expected )": 51, "expected ;": 20, "exact not last": 10, "ORE box": 9, "PR position": 7, "no-eos": 3, "expected :=": 2, "trailing tokens": 1}, "lean_wall_s": 12.343163251876831, "lean_proc_s": 110.05608749389648, "nd_verify_s": 0.15790820121765137, "workers": 96, "chunk": 400}
14
+ {"utc": "2026-10-07T20:48:17Z", "samples": 12000, "parse_fail": 1182, "distinct_checked": 10452, "both_ok": 6244, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4208, "parse_reasons": {"unbound": 872, "formula": 131, "expected (": 69, "name": 27, "exact not last": 27, "expected )": 15, "no-eos": 14, "ORE box": 10, "expected ;": 8, "PR position": 4, "trailing tokens": 3, "expected :=": 2}, "lean_wall_s": 28.084853887557983, "lean_proc_s": 725.2695417404175, "nd_verify_s": 0.5680108070373535, "workers": 96, "chunk": 400}
15
+ {"utc": "2026-10-07T20:48:41Z", "samples": 3840, "parse_fail": 757, "distinct_checked": 2863, "both_ok": 206, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2657, "parse_reasons": {"unbound": 564, "formula": 66, "expected (": 56, "name": 46, "expected )": 11, "exact not last": 7, "ORE box": 4, "no-eos": 2, "trailing tokens": 1}, "lean_wall_s": 10.690237045288086, "lean_proc_s": 76.46341276168823, "nd_verify_s": 0.1442708969116211, "workers": 96, "chunk": 400}
16
+ {"utc": "2026-10-07T20:49:10Z", "samples": 3816, "parse_fail": 897, "distinct_checked": 2679, "both_ok": 295, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2384, "parse_reasons": {"unbound": 643, "name": 79, "expected (": 71, "formula": 31, "expected )": 27, "expected :=": 16, "ORE box": 11, "exact not last": 10, "no-eos": 4, "expected =>": 2, "PR position": 1, "expected None": 1, "expected ;": 1}, "lean_wall_s": 10.29222059249878, "lean_proc_s": 69.1641194820404, "nd_verify_s": 0.14629435539245605, "workers": 96, "chunk": 400}
17
+ {"utc": "2026-10-07T20:49:48Z", "samples": 3816, "parse_fail": 985, "distinct_checked": 2632, "both_ok": 367, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2265, "parse_reasons": {"unbound": 712, "name": 128, "expected (": 87, "expected :=": 17, "exact not last": 12, "formula": 10, "expected ;": 5, "expected )": 4, "ORE box": 3, "no-eos": 3, "trailing tokens": 2, "PR position": 1, "expected :": 1}, "lean_wall_s": 10.078917264938354, "lean_proc_s": 66.81796836853027, "nd_verify_s": 0.21504807472229004, "workers": 96, "chunk": 400}
18
+ {"utc": "2026-10-07T20:51:41Z", "samples": 12000, "parse_fail": 1342, "distinct_checked": 10280, "both_ok": 6895, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3385, "parse_reasons": {"unbound": 868, "formula": 154, "expected (": 76, "name": 67, "expected ;": 42, "expected :=": 34, "expected )": 33, "exact not last": 33, "ORE box": 10, "expected fun": 7, "trailing tokens": 6, "no-eos": 6, "PR position": 6}, "lean_wall_s": 27.623512268066406, "lean_proc_s": 704.0656399726868, "nd_verify_s": 0.6165800094604492, "workers": 96, "chunk": 400}
19
+ {"utc": "2026-10-07T20:51:59Z", "samples": 3192, "parse_fail": 999, "distinct_checked": 2046, "both_ok": 155, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1891, "parse_reasons": {"unbound": 631, "name": 114, "expected (": 74, "formula": 73, "expected ;": 27, "no-eos": 20, "expected )": 18, "expected :=": 17, "expected =>": 11, "exact not last": 8, "ORE box": 3, "PR position": 2, "expected fun": 1}, "lean_wall_s": 7.663270950317383, "lean_proc_s": 38.86984157562256, "nd_verify_s": 0.1454012393951416, "workers": 96, "chunk": 400}
20
+ {"utc": "2026-10-07T20:52:24Z", "samples": 3192, "parse_fail": 976, "distinct_checked": 2054, "both_ok": 224, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1830, "parse_reasons": {"unbound": 647, "name": 111, "expected (": 84, "formula": 51, "expected )": 23, "exact not last": 20, "expected :=": 15, "expected ;": 11, "ORE box": 5, "expected fun": 4, "expected =>": 2, "no-eos": 1, "expected exact": 1, "trailing tokens": 1}, "lean_wall_s": 7.739999771118164, "lean_proc_s": 39.95041036605835, "nd_verify_s": 0.12434864044189453, "workers": 96, "chunk": 400}
21
+ {"utc": "2026-10-07T20:52:55Z", "samples": 3168, "parse_fail": 965, "distinct_checked": 2044, "both_ok": 237, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1807, "parse_reasons": {"unbound": 662, "expected (": 115, "name": 104, "formula": 37, "expected )": 13, "exact not last": 12, "expected :=": 8, "expected ;": 5, "ORE box": 4, "expected hh": 1, "PR position": 1, "expected None": 1, "expected exact": 1, "expected False": 1}, "lean_wall_s": 7.2879416942596436, "lean_proc_s": 36.48100399971008, "nd_verify_s": 0.10465478897094727, "workers": 96, "chunk": 400}
22
+ {"utc": "2026-10-07T20:54:47Z", "samples": 12000, "parse_fail": 1298, "distinct_checked": 10363, "both_ok": 7418, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2945, "parse_reasons": {"unbound": 990, "formula": 83, "expected (": 65, "expected ;": 41, "name": 36, "exact not last": 31, "expected )": 29, "ORE box": 9, "no-eos": 5, "expected :=": 4, "expected False": 2, "expected =>": 1, "trailing tokens": 1, "PR position": 1}, "lean_wall_s": 28.012216329574585, "lean_proc_s": 718.1093292236328, "nd_verify_s": 0.633544921875, "workers": 96, "chunk": 400}
23
+ {"utc": "2026-10-07T20:55:05Z", "samples": 2808, "parse_fail": 902, "distinct_checked": 1769, "both_ok": 109, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1660, "parse_reasons": {"unbound": 653, "name": 73, "expected ;": 62, "expected (": 54, "formula": 31, "expected )": 12, "ORE box": 8, "exact not last": 4, "expected :=": 3, "no-eos": 2}, "lean_wall_s": 6.543370246887207, "lean_proc_s": 30.0984787940979, "nd_verify_s": 0.0802302360534668, "workers": 96, "chunk": 400}
24
+ {"utc": "2026-10-07T20:55:26Z", "samples": 2808, "parse_fail": 882, "distinct_checked": 1775, "both_ok": 162, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1613, "parse_reasons": {"unbound": 663, "name": 81, "expected (": 62, "expected )": 24, "expected ;": 17, "formula": 14, "exact not last": 8, "PR position": 5, "ORE box": 4, "expected :=": 2, "no-eos": 2}, "lean_wall_s": 6.892028093338013, "lean_proc_s": 31.206921815872192, "nd_verify_s": 0.09554076194763184, "workers": 96, "chunk": 400}
25
+ {"utc": "2026-10-07T20:55:57Z", "samples": 2784, "parse_fail": 1017, "distinct_checked": 1653, "both_ok": 101, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1552, "parse_reasons": {"unbound": 783, "name": 95, "formula": 45, "expected (": 31, "expected )": 30, "expected :=": 12, "expected ;": 10, "PR position": 6, "ORE box": 3, "exact not last": 1, "expected .elim": 1}, "lean_wall_s": 6.296822547912598, "lean_proc_s": 25.55722737312317, "nd_verify_s": 0.08479642868041992, "workers": 96, "chunk": 400}
26
+ {"utc": "2026-10-07T20:57:52Z", "samples": 12000, "parse_fail": 1239, "distinct_checked": 10425, "both_ok": 7483, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2942, "parse_reasons": {"unbound": 904, "expected ;": 97, "formula": 75, "name": 36, "expected )": 34, "expected (": 33, "exact not last": 17, "expected :=": 15, "ORE box": 12, "no-eos": 9, "PR position": 3, "expected fun": 3, "trailing tokens": 1}, "lean_wall_s": 28.63356065750122, "lean_proc_s": 738.7091526985168, "nd_verify_s": 0.62587571144104, "workers": 96, "chunk": 400}
27
+ {"utc": "2026-10-07T20:58:08Z", "samples": 2736, "parse_fail": 894, "distinct_checked": 1703, "both_ok": 129, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1574, "parse_reasons": {"unbound": 596, "name": 131, "expected ;": 70, "expected (": 25, "no-eos": 25, "expected )": 22, "formula": 17, "exact not last": 4, "PR position": 2, "expected :=": 2}, "lean_wall_s": 6.295610189437866, "lean_proc_s": 27.089818239212036, "nd_verify_s": 0.07910776138305664, "workers": 96, "chunk": 400}
28
+ {"utc": "2026-10-07T20:58:30Z", "samples": 2712, "parse_fail": 1134, "distinct_checked": 1476, "both_ok": 215, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1261, "parse_reasons": {"unbound": 780, "name": 173, "formula": 47, "expected ;": 41, "expected (": 38, "ORE box": 16, "no-eos": 12, "expected )": 11, "exact not last": 8, "expected :=": 5, "PR position": 1, "expected fun": 1, "expected =>": 1}, "lean_wall_s": 5.237157821655273, "lean_proc_s": 19.19046711921692, "nd_verify_s": 0.08820414543151855, "workers": 96, "chunk": 400}
29
+ {"utc": "2026-10-07T20:59:06Z", "samples": 2712, "parse_fail": 1015, "distinct_checked": 1573, "both_ok": 175, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1398, "parse_reasons": {"unbound": 694, "name": 158, "expected ;": 70, "formula": 36, "expected (": 21, "expected )": 21, "ORE box": 6, "no-eos": 2, "expected :=": 2, "expected fun": 2, "exact not last": 2, "PR position": 1}, "lean_wall_s": 5.814575672149658, "lean_proc_s": 22.97928214073181, "nd_verify_s": 0.08580946922302246, "workers": 96, "chunk": 400}
30
+ {"utc": "2026-10-07T21:01:07Z", "samples": 12000, "parse_fail": 1343, "distinct_checked": 10384, "both_ok": 7714, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2670, "parse_reasons": {"unbound": 985, "expected ;": 85, "formula": 73, "expected )": 64, "expected (": 48, "name": 22, "expected :=": 20, "no-eos": 17, "exact not last": 15, "ORE box": 12, "trailing tokens": 2}, "lean_wall_s": 27.212807655334473, "lean_proc_s": 693.8677611351013, "nd_verify_s": 0.6463272571563721, "workers": 96, "chunk": 400}
31
+ {"utc": "2026-10-07T21:01:22Z", "samples": 2496, "parse_fail": 1116, "distinct_checked": 1311, "both_ok": 59, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1252, "parse_reasons": {"unbound": 771, "name": 82, "expected ;": 81, "expected (": 64, "expected )": 58, "formula": 41, "ORE box": 10, "no-eos": 6, "trailing tokens": 2, "expected :=": 1}, "lean_wall_s": 5.075615406036377, "lean_proc_s": 16.567970275878906, "nd_verify_s": 0.0572819709777832, "workers": 96, "chunk": 400}
32
+ {"utc": "2026-10-07T21:01:41Z", "samples": 2496, "parse_fail": 893, "distinct_checked": 1478, "both_ok": 90, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1388, "parse_reasons": {"unbound": 617, "name": 122, "expected ;": 50, "expected )": 28, "expected (": 24, "ORE box": 21, "formula": 20, "exact not last": 3, "PR position": 3, "no-eos": 3, "trailing tokens": 1, "expected :=": 1}, "lean_wall_s": 5.061018228530884, "lean_proc_s": 19.817551374435425, "nd_verify_s": 0.10202527046203613, "workers": 96, "chunk": 400}
33
+ {"utc": "2026-10-07T21:02:12Z", "samples": 2496, "parse_fail": 979, "distinct_checked": 1414, "both_ok": 274, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1140, "parse_reasons": {"unbound": 678, "name": 91, "expected (": 68, "expected ;": 46, "formula": 31, "no-eos": 25, "expected )": 22, "ORE box": 11, "expected :=": 3, "exact not last": 3, "PR position": 1}, "lean_wall_s": 5.075438737869263, "lean_proc_s": 19.482487678527832, "nd_verify_s": 0.10538554191589355, "workers": 96, "chunk": 400}
34
+ {"utc": "2026-10-07T21:04:14Z", "samples": 12000, "parse_fail": 1234, "distinct_checked": 10458, "both_ok": 7721, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2737, "parse_reasons": {"unbound": 884, "formula": 77, "expected ;": 65, "expected )": 46, "expected (": 46, "name": 34, "no-eos": 28, "ORE box": 16, "expected :=": 11, "exact not last": 10, "PR position": 9, "trailing tokens": 5, "expected =>": 3}, "lean_wall_s": 29.35527729988098, "lean_proc_s": 760.2542440891266, "nd_verify_s": 0.6476047039031982, "workers": 96, "chunk": 400}
35
+ {"utc": "2026-10-07T21:04:29Z", "samples": 2424, "parse_fail": 947, "distinct_checked": 1380, "both_ok": 138, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1242, "parse_reasons": {"unbound": 705, "name": 64, "expected (": 58, "formula": 41, "expected ;": 41, "no-eos": 14, "expected )": 13, "ORE box": 7, "expected :=": 2, "exact not last": 1, "expected None": 1}, "lean_wall_s": 4.973423004150391, "lean_proc_s": 18.13756513595581, "nd_verify_s": 0.08418965339660645, "workers": 96, "chunk": 400}
36
+ {"utc": "2026-10-07T21:04:48Z", "samples": 2424, "parse_fail": 944, "distinct_checked": 1371, "both_ok": 136, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1235, "parse_reasons": {"unbound": 698, "name": 87, "expected (": 66, "expected ;": 49, "no-eos": 11, "expected )": 9, "formula": 8, "PR position": 7, "ORE box": 3, "exact not last": 2, "expected :=": 1, "eof": 1, "expected =>": 1, "trailing tokens": 1}, "lean_wall_s": 4.928381443023682, "lean_proc_s": 18.114356994628906, "nd_verify_s": 0.0719904899597168, "workers": 96, "chunk": 400}
37
+ {"utc": "2026-10-07T21:05:15Z", "samples": 2424, "parse_fail": 924, "distinct_checked": 1390, "both_ok": 204, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1186, "parse_reasons": {"unbound": 705, "name": 98, "expected (": 36, "expected )": 30, "expected :=": 20, "expected ;": 16, "formula": 9, "exact not last": 3, "no-eos": 3, "PR position": 2, "expected :": 1, "expected =>": 1}, "lean_wall_s": 5.1239118576049805, "lean_proc_s": 18.735954761505127, "nd_verify_s": 0.07892227172851562, "workers": 96, "chunk": 400}
38
+ {"utc": "2026-10-07T21:07:17Z", "samples": 12000, "parse_fail": 1181, "distinct_checked": 10514, "both_ok": 8108, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2406, "parse_reasons": {"unbound": 910, "formula": 64, "name": 42, "expected :=": 38, "expected )": 33, "expected (": 33, "expected ;": 29, "ORE box": 14, "no-eos": 9, "exact not last": 7, "PR position": 1, "expected =>": 1}, "lean_wall_s": 28.422096014022827, "lean_proc_s": 737.6269717216492, "nd_verify_s": 0.6421933174133301, "workers": 96, "chunk": 400}
39
+ {"utc": "2026-10-07T21:07:32Z", "samples": 2208, "parse_fail": 801, "distinct_checked": 1317, "both_ok": 97, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1220, "parse_reasons": {"unbound": 641, "name": 66, "expected ;": 32, "expected )": 15, "formula": 15, "ORE box": 12, "expected (": 9, "exact not last": 6, "no-eos": 2, "PR position": 1, "expected =>": 1, "expected :=": 1}, "lean_wall_s": 4.930431842803955, "lean_proc_s": 16.647331476211548, "nd_verify_s": 0.06029629707336426, "workers": 96, "chunk": 400}
40
+ {"utc": "2026-10-07T21:07:50Z", "samples": 2208, "parse_fail": 830, "distinct_checked": 1283, "both_ok": 46, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1237, "parse_reasons": {"unbound": 601, "name": 106, "expected (": 30, "expected ;": 28, "ORE box": 28, "expected )": 23, "no-eos": 4, "formula": 3, "exact not last": 3, "expected :=": 3, "PR position": 1}, "lean_wall_s": 4.524183511734009, "lean_proc_s": 13.881052494049072, "nd_verify_s": 0.05957388877868652, "workers": 96, "chunk": 400}
41
+ {"utc": "2026-10-07T21:08:16Z", "samples": 2184, "parse_fail": 765, "distinct_checked": 1330, "both_ok": 142, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1188, "parse_reasons": {"unbound": 537, "name": 114, "expected ;": 52, "expected (": 34, "ORE box": 8, "expected )": 7, "PR position": 5, "exact not last": 4, "formula": 3, "expected :=": 1}, "lean_wall_s": 4.657480716705322, "lean_proc_s": 15.903783321380615, "nd_verify_s": 0.09911918640136719, "workers": 96, "chunk": 400}
42
+ {"utc": "2026-10-07T21:10:18Z", "samples": 12000, "parse_fail": 1178, "distinct_checked": 10530, "both_ok": 8176, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2354, "parse_reasons": {"unbound": 847, "expected (": 87, "formula": 79, "name": 46, "expected ;": 24, "expected :=": 22, "ORE box": 16, "no-eos": 15, "exact not last": 14, "trailing tokens": 12, "expected )": 11, "PR position": 4, "expected exact": 1}, "lean_wall_s": 28.515458345413208, "lean_proc_s": 743.6862316131592, "nd_verify_s": 0.6583049297332764, "workers": 96, "chunk": 400}
43
+ {"utc": "2026-10-07T21:10:31Z", "samples": 2112, "parse_fail": 803, "distinct_checked": 1217, "both_ok": 117, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1100, "parse_reasons": {"unbound": 590, "expected (": 84, "name": 62, "ORE box": 19, "expected ;": 14, "expected )": 13, "formula": 12, "exact not last": 5, "expected :=": 2, "no-eos": 1, "PR position": 1}, "lean_wall_s": 4.321662902832031, "lean_proc_s": 13.072744131088257, "nd_verify_s": 0.06403446197509766, "workers": 96, "chunk": 400}
44
+ {"utc": "2026-10-07T21:10:50Z", "samples": 2112, "parse_fail": 944, "distinct_checked": 1094, "both_ok": 93, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1001, "parse_reasons": {"unbound": 645, "name": 172, "expected (": 42, "formula": 35, "expected :=": 16, "ORE box": 12, "expected )": 8, "PR position": 7, "exact not last": 3, "expected ;": 3, "eof": 1}, "lean_wall_s": 4.011633396148682, "lean_proc_s": 11.357386589050293, "nd_verify_s": 0.049758195877075195, "workers": 96, "chunk": 400}
45
+ {"utc": "2026-10-07T21:11:22Z", "samples": 2088, "parse_fail": 892, "distinct_checked": 1108, "both_ok": 143, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 965, "parse_reasons": {"unbound": 590, "name": 154, "expected (": 39, "expected )": 30, "formula": 29, "expected :=": 24, "expected ;": 13, "no-eos": 5, "ORE box": 4, "exact not last": 2, "PR position": 2}, "lean_wall_s": 4.066845655441284, "lean_proc_s": 11.552957534790039, "nd_verify_s": 0.05251193046569824, "workers": 96, "chunk": 400}
46
+ {"utc": "2026-10-07T21:25:24Z", "samples": 70912, "parse_fail": 15140, "distinct_checked": 45908, "both_ok": 27000, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 18907, "parse_reasons": {"unbound": 10449, "name": 2652, "expected (": 436, "formula": 397, "expected ;": 278, "expected )": 277, "ORE box": 234, "no-eos": 227, "PR position": 108, "exact not last": 62, "expected :=": 6, "eof": 4, "expected None": 3, "expected exact": 2, "trailing tokens": 2, "expected :": 1, "expected fun": 1, "expected .elim": 1}, "lean_wall_s": 107.77420139312744, "lean_proc_s": 8648.88804268837, "nd_verify_s": 2.7642552852630615, "workers": 96, "chunk": 400}
47
+ {"utc": "2026-10-07T21:25:38Z", "samples": 2000, "parse_fail": 5, "distinct_checked": 1995, "both_ok": 1975, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 20, "parse_reasons": {"unbound": 3, "expected (": 1, "name": 1}, "lean_wall_s": 3.406017303466797, "lean_proc_s": 15.83350419998169, "nd_verify_s": 0.09989452362060547, "workers": 96, "chunk": 400}
48
+ {"utc": "2026-10-07T21:31:22Z", "samples": 70912, "parse_fail": 15479, "distinct_checked": 46015, "both_ok": 24963, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 21052, "parse_reasons": {"unbound": 10622, "name": 2211, "expected (": 651, "formula": 605, "expected )": 402, "expected ;": 388, "ORE box": 245, "no-eos": 214, "PR position": 62, "exact not last": 61, "trailing tokens": 5, "expected :": 3, "expected None": 2, "expected =>": 2, "eof": 2, "expected .elim": 1, "expected exact": 1, "expected fun": 1, "expected :=": 1}, "lean_wall_s": 113.09152698516846, "lean_proc_s": 8962.125233411789, "nd_verify_s": 2.718319892883301, "workers": 96, "chunk": 400}
runs/507-fable_ft-token-batches_s1/pretrain.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99c053830f4f5858de5fbaee8cfbdd74dc73478a4d802477abd0af048e4ba6fd
3
+ size 25625153
runs/507-fable_ft-token-batches_s1/source/harness.py ADDED
@@ -0,0 +1,529 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ # imports the fork (model, tokenizer, train, ...), which is not resolvable here
3
+ # pyright: reportMissingImports=false
4
+ """FROZEN harness for the test-time-loops autoresearch loop. The research agent does not edit it.
5
+
6
+ Forked from ../autoresearch/harness.py (sha 2aa52f23e7ec) on 2026-10-07. What changed: the RL
7
+ stage belongs to the agent, behind a wall-clock budget and a sampling oracle, and the metric is
8
+ measured with the model's test-time compute knob (KNOB) set to EVAL_LEVEL.
9
+
10
+ One timed run = harness owns the clocks, the oracle and the eval; `pretrain.py` (agent-editable)
11
+ owns pretraining and RL:
12
+
13
+ 1. load frozen data (Dan's cap-6 Stage-1 file, ladder RL-target pool, ladder transfer pool)
14
+ 2. call pretrain.pretrain(train_recs, PRETRAIN_SECS, ...) -> (model, tok, stats); charged for
15
+ everything including model construction and torch.compile. Overrun past the grace -> timeout.
16
+ 3. check n_params <= PARAM_CAP, tok.mode == TOK_MODE and that the model has the KNOB attribute
17
+ 4. call pretrain.rl(model, tok, targets, train_recs, oracle, RL_SECS, ...): expert iteration in
18
+ whatever form the agent likes. The targets are statements only, and `oracle.sample` is the
19
+ only way to turn them into training data: it samples the model's own attempts and returns
20
+ the ones Lean AND nd_verify accept. RL_SECS is wall-clock inside rl() minus the time the
21
+ oracle spent in the Lean gate (pod CPUs differ; GPU work is what is rationed). Overrun past
22
+ the grace, or an oracle call after the budget is spent -> timeout.
23
+ 5. eval on the dev half of the transfer pool with the harness setting the model's KNOB:
24
+ METRIC = dev theorems solved at L_true >= 7 with EVAL_K attempts at EVAL_LEVEL.
25
+ The same eval at DIAG_LEVEL is recorded as a diagnostic, not scored.
26
+ 6. SIGALRM at TOTAL_SECS from process start kills everything -> status "timeout"
27
+
28
+ The holdout half of the transfer pool is never scored here. Everything is written to --out.
29
+
30
+ python3 harness.py --out artifacts/x # on the pod (see pod.py / remote.sh)
31
+ ND_FORK=... python3 harness.py --out /tmp/smoke --smoke # CPU, ~1 min, plumbing check
32
+ """
33
+
34
+ import argparse
35
+ import collections
36
+ import hashlib
37
+ import importlib
38
+ import json
39
+ import os
40
+ import random
41
+ import shutil
42
+ import signal
43
+ import sys
44
+ import time
45
+ import traceback
46
+ from pathlib import Path
47
+
48
+ HERE = os.path.dirname(os.path.abspath(__file__))
49
+ FORK = os.environ.get("ND_FORK", "/workspace/nd-takehome")
50
+ sys.path.insert(0, HERE)
51
+ sys.path.insert(
52
+ 0, FORK
53
+ ) # fork first: its train.py provides batch/loss_on; the agent's file is pretrain.py
54
+ import torch # noqa: E402
55
+ from eval_set import judge, wilson # noqa: E402
56
+ from sample import generate # noqa: E402
57
+
58
+ # ---------------------------------------------------------------- the rules
59
+ # (hash-checked into every result)
60
+ TOTAL_SECS = int(
61
+ os.environ.get("ND_TOTAL_SECS", 5400)
62
+ ) # process start -> summary written; SIGALRM past this. env: benchmarks only
63
+ PRETRAIN_SECS = 300 # agent's pretraining budget, compile included
64
+ PRETRAIN_GRACE = 15 # seconds the agent may overrun before the run is a "timeout"
65
+ RL_SECS = 1800 # agent's RL budget: wall-clock inside rl() minus the oracle's Lean-gate time
66
+ RL_GRACE = 60
67
+ PARAM_CAP = 10_000_000 # unique trainable params, embeddings included; checked after RL too
68
+ N_TARGETS, TEMP, MAX_NEW, SAMPLE_BATCH = 1500, 0.8, 512, 1024
69
+ EVAL_K = 64
70
+ # The model's test-time compute knob: an integer attribute the harness sets before each eval and
71
+ # rl() sets before oracle calls. A sibling loop for another architecture changes these three
72
+ # constants and nothing else in this file.
73
+ KNOB = "n_loops" # loops of the shared core
74
+ EVAL_LEVEL = 8 # the metric: the knob's value during the scored eval
75
+ DIAG_LEVEL = 2 # the same eval again at this value, recorded but not scored
76
+ TOK_MODE = "lean_seq" # required tokenizer mode (fork lean_tok.py)
77
+ DEV = "cuda" if torch.cuda.is_available() else "cpu"
78
+
79
+
80
+ def read(fn):
81
+ return [json.loads(line) for line in open(fn) if line.strip()]
82
+
83
+
84
+ def log(msg):
85
+ print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
86
+
87
+
88
+ def sha(fn):
89
+ return hashlib.sha256(open(fn, "rb").read()).hexdigest()[:12]
90
+
91
+
92
+ class Timeout(Exception):
93
+ pass
94
+
95
+
96
+ class BudgetSpent(Timeout):
97
+ """rl() asked the oracle for samples after RL_SECS was used up."""
98
+
99
+
100
+ # ---------------------------------------------------------------- bookkeeping
101
+
102
+
103
+ def save_checkpoint(a, model, tok, stage):
104
+ """Publish complete weight snapshots atomically; these are not optimizer-resume files."""
105
+ path = Path(a.out) / f"{stage}.pt"
106
+ pending = path.with_suffix(".pt.tmp")
107
+ try:
108
+ torch.save(
109
+ {
110
+ "state": model.state_dict(),
111
+ "tok_mode": getattr(tok, "mode", None),
112
+ "stage": stage,
113
+ "seed": a.seed,
114
+ "harness_sha": sha(f"{HERE}/harness.py"),
115
+ "pretrain_sha": sha(f"{HERE}/pretrain.py"),
116
+ },
117
+ pending,
118
+ )
119
+ os.replace(pending, path)
120
+ finally:
121
+ pending.unlink(missing_ok=True)
122
+
123
+
124
+ def lstar(solved_names, recs, need=5):
125
+ ls_solved = [r["n_lines"] for r in recs if r["name"] in solved_names]
126
+ ge = {L: sum(1 for x in ls_solved if x >= L) for L in range(7, 15)}
127
+ return max([L for L, c in ge.items() if c >= need], default=0), ge
128
+
129
+
130
+ def by_bin(solved_names, recs):
131
+ by = collections.defaultdict(lambda: [0, 0])
132
+ for r in recs:
133
+ by[r["n_lines"]][0] += r["name"] in solved_names
134
+ by[r["n_lines"]][1] += 1
135
+ return {
136
+ str(L): {"solved": k, "n": n, "rate": k / n, "ci": list(wilson(k, n))}
137
+ for L, (k, n) in sorted(by.items())
138
+ }
139
+
140
+
141
+ def gate_seconds():
142
+ """Seconds the fork's lean_gate has logged so far (Lean wall-clock plus nd_verify)."""
143
+ fn = os.environ.get("LEAN_GATE_LOG")
144
+ if not fn or not os.path.exists(fn):
145
+ return 0.0
146
+ return sum(rec["lean_wall_s"] + rec["nd_verify_s"] for rec in read(fn))
147
+
148
+
149
+ def sample_and_judge(model, tok, recs, k, temperature, seed, batch, max_new):
150
+ prompts = [r["prompt"] for r in recs for _ in range(k)]
151
+ t0, g0 = time.time(), gate_seconds()
152
+ flat = generate(
153
+ model,
154
+ tok,
155
+ prompts,
156
+ greedy=(k == 1 and temperature == 0),
157
+ temperature=temperature,
158
+ batch=batch,
159
+ seed=seed,
160
+ max_new=max_new,
161
+ )
162
+ secs, gate = time.time() - t0, gate_seconds() - g0
163
+ rows = judge(recs, [flat[i * k : (i + 1) * k] for i in range(len(recs))], "n_lines")
164
+ return rows, {
165
+ "samples": len(prompts),
166
+ "secs": secs,
167
+ "gate_secs": gate,
168
+ "samples_per_sec": len(prompts) / secs,
169
+ }
170
+
171
+
172
+ def greedy_rate(model, tok, recs):
173
+ rows, _ = sample_and_judge(model, tok, recs, 1, 0, 0, SAMPLE_BATCH, MAX_NEW)
174
+ return sum(r["solved"] for r in rows) / max(len(rows), 1)
175
+
176
+
177
+ def dev_split(transfer):
178
+ return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 0]
179
+
180
+
181
+ def holdout_split(transfer):
182
+ """Never scored by the loop. Score final.pt on this offline when checking in."""
183
+ return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 1]
184
+
185
+
186
+ # ---------------------------------------------------------------- the RL oracle
187
+
188
+
189
+ class Oracle:
190
+ """What rl() gets instead of data: the model's own attempts at the RL targets, filtered to
191
+ the ones Lean AND nd_verify accept. It also keeps the RL clock and the record of every call.
192
+ """
193
+
194
+ def __init__(self, a, tok, targets):
195
+ self.a, self.tok = a, tok
196
+ self.targets = {t["name"]: t for t in targets}
197
+ self.found = collections.defaultdict(dict) # name -> {proof text: record}
198
+ self.calls = []
199
+ self.t0 = self.last = time.time()
200
+ self.uncharged = 0.0
201
+
202
+ def used(self):
203
+ """RL seconds charged so far: wall-clock since rl() started, minus Lean-gate time."""
204
+ return time.time() - self.t0 - self.uncharged
205
+
206
+ def left(self):
207
+ return self.a.rl_secs - self.used()
208
+
209
+ def sample(self, model, targets, k, temperature=TEMP, batch=SAMPLE_BATCH):
210
+ """k attempts at each of `targets` (records rl() was given, any subset, in any order) from
211
+ `model` as it is now, in eval mode, at whatever its KNOB is set to. Returns one
212
+ dict per target, in order: name, prompt, n_lines, proofs (the distinct accepted proof
213
+ texts), n_ok (accepted attempts) and n_tried. Rejected attempts are not returned.
214
+ Raises BudgetSpent, which ends the run as a timeout, once the budget is used up."""
215
+ if self.left() <= 0:
216
+ raise BudgetSpent(f"oracle called with {self.used():.0f}s of {self.a.rl_secs}s used")
217
+ recs = [self.targets[t["name"]] for t in targets]
218
+ call = len(self.calls) + 1
219
+ model.eval()
220
+ rows, samp = sample_and_judge(
221
+ model, self.tok, recs, k, temperature, self.a.seed * 1000 + call, batch, self.a.max_new
222
+ )
223
+ self.uncharged += samp["gate_secs"]
224
+ new = 0
225
+ for t, row in zip(recs, rows, strict=True):
226
+ for p, wl, pl in zip(
227
+ row["proofs"], row["written_lens"], row["pruned_lens"], strict=True
228
+ ):
229
+ if p not in self.found[t["name"]]:
230
+ self.found[t["name"]][p] = {
231
+ "proof": p,
232
+ "written": wl,
233
+ "pruned": pl,
234
+ "round": call,
235
+ }
236
+ new += 1
237
+ solved = {n for n, d in self.found.items() if d}
238
+ ls, ge = lstar(solved, list(self.targets.values()))
239
+ now = time.time()
240
+ self.calls.append(
241
+ {
242
+ "round": call,
243
+ "level": getattr(model, KNOB, None),
244
+ "k": k,
245
+ "n_targets": len(recs),
246
+ "sample_acc": sum(x["n_ok"] for x in rows) / max(samp["samples"], 1),
247
+ "new_proofs": new,
248
+ "targets_solved": len(solved),
249
+ "lstar": ls,
250
+ "ge": ge,
251
+ "mix_rl_records": sum(len(x["proofs"]) for x in rows), # proofs handed back
252
+ "sampling": samp,
253
+ "secs": now - self.last, # since the previous call returned: includes rl()'s work
254
+ "rl_secs_used": self.used(),
255
+ }
256
+ )
257
+ self.last = now
258
+ c = self.calls[-1]
259
+ log(
260
+ f"ei r{call}: acc {c['sample_acc']:.4f}, +{new} proofs, "
261
+ f"solved {len(solved)}/{len(self.targets)}, L*={ls}, {KNOB} {c['level']}, "
262
+ f"{len(recs)} targets x {k}, {samp['samples_per_sec']:.0f} samples/s, "
263
+ f"RL clock {c['rl_secs_used']:.0f}/{self.a.rl_secs}s"
264
+ )
265
+ return [
266
+ {
267
+ "name": t["name"],
268
+ "prompt": t["prompt"],
269
+ "n_lines": t["n_lines"],
270
+ "proofs": list(row["proofs"]),
271
+ "n_ok": row["n_ok"],
272
+ "n_tried": row["n_tried"],
273
+ }
274
+ for t, row in zip(recs, rows, strict=True)
275
+ ]
276
+
277
+
278
+ # ---------------------------------------------------------------- stages
279
+
280
+
281
+ def stage_rl(a, agent, tok, model, targets, train_recs):
282
+ oracle = Oracle(a, tok, targets)
283
+ # statements only: what the agent's RL may know about a target
284
+ visible = [{"name": t["name"], "prompt": t["prompt"], "n_lines": t["n_lines"]} for t in targets]
285
+ stats = agent.rl(model, tok, visible, train_recs, oracle, a.rl_secs, DEV, a.seed, log)
286
+ if DEV == "cuda":
287
+ torch.cuda.synchronize()
288
+ used = oracle.used()
289
+ model.eval()
290
+ with open(f"{a.out}/found_targets.jsonl", "w") as f:
291
+ for t in targets:
292
+ for x in oracle.found[t["name"]].values():
293
+ f.write(
294
+ json.dumps({"name": t["name"], "thm": t["thm"], "L_true": t["n_lines"], **x})
295
+ + "\n"
296
+ )
297
+ log(f"rl returned: {used:.0f}s of {a.rl_secs}s charged, {oracle.uncharged:.0f}s Lean gate")
298
+ return {
299
+ "rounds": oracle.calls,
300
+ "secs": time.time() - oracle.t0,
301
+ "rl_secs_used": used,
302
+ "gate_secs": oracle.uncharged,
303
+ "targets": len(targets),
304
+ "stats": dict(stats or {}),
305
+ }
306
+
307
+
308
+ def eval_at(a, tok, model, dev_pool, level):
309
+ setattr(model, KNOB, level)
310
+ rows, samp = sample_and_judge(
311
+ model, tok, dev_pool, a.eval_k, TEMP, a.seed * 1000 + 777, SAMPLE_BATCH, a.max_new
312
+ )
313
+ solved = {r["name"] for r in rows if r["solved"]}
314
+ ls, ge = lstar(solved, dev_pool)
315
+ out = {
316
+ "level": level,
317
+ "attempts_per_theorem": a.eval_k,
318
+ "solved": len(solved),
319
+ "n": len(dev_pool),
320
+ "solved_ge7": ge[7],
321
+ "lstar": ls,
322
+ "ge": ge,
323
+ "by_bin": by_bin(solved, dev_pool),
324
+ "sample_acc": sum(r["n_ok"] for r in rows) / max(samp["samples"], 1),
325
+ "max_pruned": max((max(r["pruned_lens"]) for r in rows if r["solved"]), default=0),
326
+ "sampling": samp,
327
+ }
328
+ log(
329
+ f"eval at {KNOB} {level}: solved {len(solved)}/{len(dev_pool)}, ge7 {ge[7]}, L*={ls}, "
330
+ f"per-sample {out['sample_acc']:.4f}, {samp['samples_per_sec']:.0f} samples/s"
331
+ )
332
+ return out, rows
333
+
334
+
335
+ def stage_eval(a, tok, model, dev_pool, heldout_recs):
336
+ out, rows = eval_at(a, tok, model, dev_pool, a.eval_level)
337
+ out["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
338
+ with open(f"{a.out}/eval_dev.jsonl", "w") as f:
339
+ for r in rows:
340
+ keys = ("name", "n_lines", "solved", "n_ok", "proofs", "written_lens", "pruned_lens")
341
+ f.write(json.dumps({k: r[k] for k in keys}) + "\n")
342
+ diag, _ = eval_at(a, tok, model, dev_pool, a.diag_level)
343
+ out["diag"] = diag
344
+ setattr(model, KNOB, a.eval_level)
345
+ log(
346
+ f"eval: solved {out['solved']}/{len(dev_pool)}, ge7 {out['solved_ge7']} at "
347
+ f"{KNOB} {a.eval_level} ({diag['solved_ge7']} at {a.diag_level}), L*={out['lstar']}, "
348
+ f"max pruned {out['max_pruned']}, held-out greedy {out['heldout_greedy']:.3f}"
349
+ )
350
+ return out
351
+
352
+
353
+ def lean_gate_totals():
354
+ """Sum the fork lean_gate's per-generate() log lines (remote.sh sets LEAN_GATE_LOG)."""
355
+ fn = os.environ.get("LEAN_GATE_LOG")
356
+ if not fn or not os.path.exists(fn):
357
+ return None
358
+ keys = ("samples", "parse_fail", "distinct_checked", "both_ok", "nd_ok_lean_rej")
359
+ tot = dict.fromkeys(keys + ("nd_rej_lean_ok", "lean_wall_s"), 0)
360
+ for rec in read(fn):
361
+ for k in tot:
362
+ tot[k] += rec[k]
363
+ return tot
364
+
365
+
366
+ RULES = (
367
+ "TOTAL_SECS",
368
+ "PRETRAIN_SECS",
369
+ "RL_SECS",
370
+ "PARAM_CAP",
371
+ "N_TARGETS",
372
+ "EVAL_K",
373
+ "KNOB",
374
+ "EVAL_LEVEL",
375
+ "DIAG_LEVEL",
376
+ "TOK_MODE",
377
+ )
378
+
379
+
380
+ def write_summary(a, status, **kw):
381
+ s = {
382
+ "status": status,
383
+ "metric": kw.get("eval", {}).get("solved_ge7"),
384
+ "metric_diag": kw.get("eval", {}).get("diag", {}).get("solved_ge7"),
385
+ "metric_rl": (kw.get("rl", {}).get("rounds") or [{}])[-1].get(
386
+ "targets_solved"
387
+ ), # secondary: cumulative RL targets solved
388
+ "harness_sha": sha(f"{HERE}/harness.py"),
389
+ "pretrain_sha": sha(f"{HERE}/pretrain.py"),
390
+ "device": DEV,
391
+ "gpu": torch.cuda.get_device_name(0) if DEV == "cuda" else "cpu",
392
+ "torch": torch.__version__,
393
+ "lean_gate": lean_gate_totals(),
394
+ "total_secs": time.time() - a.T0,
395
+ "rules": {k: globals()[k] for k in RULES},
396
+ **kw,
397
+ }
398
+ with open(f"{a.out}/summary.json", "w") as f:
399
+ json.dump(s, f, indent=1)
400
+ log(
401
+ f"SUMMARY status={status} metric={s['metric']} metric_diag={s['metric_diag']} "
402
+ f"metric_rl={s['metric_rl']} total {s['total_secs']:.0f}s"
403
+ )
404
+ return s
405
+
406
+
407
+ def main():
408
+ ap = argparse.ArgumentParser()
409
+ ap.add_argument("--out", required=True)
410
+ ap.add_argument("--seed", type=int, default=0)
411
+ ap.add_argument(
412
+ "--pretrain-only", action="store_true", help="save pretrain.pt and stop before eval/RL"
413
+ )
414
+ ap.add_argument(
415
+ "--smoke", action="store_true", help="tiny everything, for a CPU plumbing check"
416
+ )
417
+ a = ap.parse_args()
418
+ a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = (
419
+ PRETRAIN_SECS,
420
+ RL_SECS,
421
+ EVAL_K,
422
+ MAX_NEW,
423
+ N_TARGETS,
424
+ )
425
+ a.eval_level, a.diag_level = EVAL_LEVEL, DIAG_LEVEL
426
+ if a.smoke:
427
+ a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = 10, 30, 2, 120, 8
428
+ os.makedirs(a.out, exist_ok=True)
429
+ # Preserve discarded variants too: weights alone do not specify the model architecture.
430
+ source_dir = Path(a.out) / "source"
431
+ source_dir.mkdir(exist_ok=True)
432
+ for filename in ("pretrain.py", "harness.py", "remote.sh"):
433
+ shutil.copy2(Path(HERE) / filename, source_dir / filename)
434
+ a.T0 = time.time()
435
+ torch.manual_seed(a.seed)
436
+
437
+ def on_alarm(signum, frame):
438
+ raise Timeout()
439
+
440
+ signal.signal(signal.SIGALRM, on_alarm)
441
+ signal.alarm(TOTAL_SECS if not a.smoke else 900)
442
+
443
+ train_recs = read(f"{FORK}/data/train.jsonl")
444
+ assert all(r["n_lines"] <= 6 for r in train_recs), "pretrain file violates cap 6"
445
+ heldout_all = read(f"{FORK}/data/heldout.jsonl")
446
+ # Length-ordered file: take 2000 evenly spaced records, so every length 2-6 is covered.
447
+ heldout_all = [heldout_all[i * len(heldout_all) // 2000] for i in range(2000)]
448
+ heldout_recs = heldout_all if not a.smoke else heldout_all[:16]
449
+ targets_all = read(f"{FORK}/data/ladder/rl_targets.jsonl")
450
+ transfer = read(f"{FORK}/data/ladder/transfer.jsonl")
451
+ dev_pool = dev_split(transfer) if not a.smoke else dev_split(transfer)[:8]
452
+ targets = random.Random(a.seed).sample(targets_all, min(a.n_targets, len(targets_all)))
453
+ tk, dk, hk = (
454
+ {t["key"] for t in targets},
455
+ {r["key"] for r in dev_pool},
456
+ {r["key"] for r in train_recs},
457
+ )
458
+ assert not (tk & dk) and not (tk & hk) and not (dk & hk), "renaming-class leak between pools"
459
+ log(
460
+ f"pools: train {len(train_recs):,}, targets {len(targets)}, dev-transfer {len(dev_pool)}, "
461
+ f"held-out {len(heldout_recs)}; "
462
+ f"device {DEV}; harness {sha(f'{HERE}/harness.py')} pretrain {sha(f'{HERE}/pretrain.py')}"
463
+ )
464
+
465
+ pre = {}
466
+ try:
467
+ agent = importlib.import_module("pretrain")
468
+ t0 = time.time()
469
+ model, tok, pre = agent.pretrain(train_recs, a.pretrain_secs, DEV, a.seed, log)
470
+ if DEV == "cuda":
471
+ torch.cuda.synchronize()
472
+ pre = dict(pre or {})
473
+ pre["secs"] = time.time() - t0
474
+ model.eval()
475
+ n_params = sum(p.numel() for p in model.parameters())
476
+ pre["n_params"] = n_params
477
+ log(
478
+ f"pretrain returned after {pre['secs']:.0f}s (budget {a.pretrain_secs}s), "
479
+ f"{n_params:,} params"
480
+ )
481
+ if pre["secs"] > a.pretrain_secs + PRETRAIN_GRACE and not a.smoke:
482
+ over = pre["secs"] - a.pretrain_secs
483
+ write_summary(
484
+ a, "timeout", pretrain=pre, reason=f"pretrain overran budget by {over:.0f}s"
485
+ )
486
+ return
487
+ if getattr(tok, "mode", None) != TOK_MODE:
488
+ write_summary(a, "crash", pretrain=pre, reason=f"tok.mode {tok.mode!r} != {TOK_MODE!r}")
489
+ return
490
+ if n_params > PARAM_CAP:
491
+ write_summary(a, "crash", pretrain=pre, reason=f"{n_params} params > cap {PARAM_CAP}")
492
+ return
493
+ if not isinstance(getattr(model, KNOB, None), int):
494
+ reason = f"model has no integer {KNOB} for the harness to set at eval"
495
+ write_summary(a, "crash", pretrain=pre, reason=reason)
496
+ return
497
+ save_checkpoint(a, model, tok, "pretrain")
498
+ if a.pretrain_only:
499
+ write_summary(a, "ok", stage="pretrain_only", pretrain=pre)
500
+ return
501
+ pre["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
502
+ log(f"pretrain held-out greedy {pre['heldout_greedy']:.3f}")
503
+ rl = stage_rl(a, agent, tok, model, targets, train_recs)
504
+ if rl["rl_secs_used"] > a.rl_secs + RL_GRACE and not a.smoke:
505
+ over = rl["rl_secs_used"] - a.rl_secs
506
+ reason = f"rl overran budget by {over:.0f}s"
507
+ write_summary(a, "timeout", pretrain=pre, rl=rl, reason=reason)
508
+ return
509
+ n_after = sum(p.numel() for p in model.parameters())
510
+ if n_after > PARAM_CAP:
511
+ reason = f"{n_after} params > cap {PARAM_CAP} after rl"
512
+ write_summary(a, "crash", pretrain=pre, rl=rl, reason=reason)
513
+ return
514
+ save_checkpoint(a, model, tok, "final")
515
+ ev = stage_eval(a, tok, model, dev_pool, heldout_recs)
516
+ write_summary(a, "ok", pretrain=pre, rl=rl, eval=ev)
517
+ except BudgetSpent as e:
518
+ write_summary(a, "timeout", pretrain=pre, reason=f"rl: {e}")
519
+ except Timeout:
520
+ write_summary(a, "timeout", pretrain=pre, reason=f"exceeded TOTAL_SECS={TOTAL_SECS}")
521
+ except Exception:
522
+ traceback.print_exc()
523
+ write_summary(a, "crash", pretrain=pre, reason=traceback.format_exc()[-2000:])
524
+ finally:
525
+ signal.alarm(0)
526
+
527
+
528
+ if __name__ == "__main__":
529
+ main()
runs/507-fable_ft-token-batches_s1/source/pretrain.py ADDED
@@ -0,0 +1,585 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # imports the fork (model, tokenizer, train, ...), which is not resolvable here
2
+ # pyright: reportMissingImports=false
3
+ """AGENT-EDITABLE: pretraining and RL. This is the only file the autoresearch agent changes.
4
+
5
+ Contract with harness.py (frozen):
6
+
7
+ pretrain(train_recs, budget_secs, device, seed, log) -> (model, tok, stats)
8
+
9
+ train_recs list of dicts {prompt, proof, n_lines<=6, key, ...}; the frozen cap-6 Stage-1 data
10
+ budget_secs wall-clock you are charged for, from call to return, INCLUDING model construction
11
+ and torch.compile. Overrun by more than the harness grace -> the run is a timeout.
12
+ model nn.Module; <= PARAM_CAP unique params (checked again after rl). The harness samples
13
+ from it with the fork's sample.generate, so it must provide:
14
+ - forward(idx) -> logits (B,T,V) and forward(idx, pos=, mask=, caches=) with a
15
+ per-block KV cache dict, exactly like the fork's GPT (model.blocks must exist)
16
+ - an integer `n_loops` (the harness's KNOB): the loop count applied in eval
17
+ mode. The harness sets it to EVAL_LEVEL for the scored eval and to DIAG_LEVEL
18
+ for the diagnostic one.
19
+ tok must provide encode_prompt, encode_proof, decode, shift_abs, pad, eos, mode, and
20
+ mode must be "lean_seq" (harness TOK_MODE): Lean text in and out, Lean-gated samples.
21
+ Decoder rule: decode() is deterministic and single-pass; it may not call the
22
+ verifier or try alternatives.
23
+ stats dict recorded in summary.json and copied into results.json. MUST contain `curve`:
24
+ a list of {secs, step, train_loss, val_loss} taken every ~CURVE_EVERY seconds
25
+ (see log_point below), so a run's trajectory survives in git. Add anything else.
26
+ val_loss is on VAL_N records held out from train_recs by this file; keep that.
27
+
28
+ rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log) -> stats | None
29
+
30
+ targets list of dicts {name, prompt, n_lines}: the 1,500 RL target statements. No proofs.
31
+ oracle oracle.sample(model, targets_subset, k, temperature=0.8, batch=1024) samples k
32
+ attempts per target from the model in eval mode at its current n_loops, and returns
33
+ per target {name, prompt, n_lines, proofs, n_ok, n_tried}; `proofs` are the distinct
34
+ attempts Lean AND nd_verify accepted. This is the only source of new training data.
35
+ oracle.used() / oracle.left(): seconds of the RL budget charged / remaining.
36
+ budget_secs RL_SECS: wall-clock inside rl() minus the oracle's Lean-gate time. Returning more
37
+ than the grace late, or calling the oracle with nothing left, is a timeout.
38
+ The model is trained in place. Everything else about RL is yours: rounds, samples per target,
39
+ loop counts when sampling, which proofs to keep, weighting, replay, optimizer, loss.
40
+
41
+ Anything imported from the fork (ND_FORK) is fair game to copy in here and modify.
42
+ """
43
+
44
+ import math
45
+ import random
46
+ import time
47
+
48
+ import torch
49
+ import torch._inductor.config # noqa: F401 (explicit, so the attribute resolves for pyright)
50
+ from lean_tok import MAXN # fork
51
+ from model import GPT, Block # fork
52
+ from normalize import norm # fork
53
+ from tokenizer import make_tokenizer # fork
54
+ from train import loss_on # fork
55
+
56
+ # ---- knobs
57
+ # Unique blocks: one untied entry block, N_CORE shared core blocks, one untied exit block.
58
+ N_CORE = 2
59
+ N_LAYER, D, N_HEAD = 2 + N_CORE, 384, 8
60
+ # Core applications (loops). Training draws one count per batch from TRAIN_LOOPS, in pretraining
61
+ # and in rl()'s fine-tune; eval and sampling apply model.n_loops, which the harness sets for its
62
+ # evals and rl() sets before each oracle call. DEFAULT_LOOPS is used for the validation curve and
63
+ # the harness's post-pretraining greedy check.
64
+ TRAIN_LOOPS = (1, 2, 3, 4, 5, 6, 7, 8)
65
+ DEFAULT_LOOPS = 2
66
+ VAL_LOOPS = (1, 2, 4, 8) # val loss at each count, once, when pretraining ends
67
+ # ---- RL knobs: the old harness's expert iteration (600 AdamW steps per round on up to 4 proofs
68
+ # per solved target, weighted x4, mixed with 20,000 replayed pretraining records), except for
69
+ # where the samples go. Each round every target gets RL_K attempts at RL_FIRST_LOOPS; a target
70
+ # all of them fail on gets RL_RETRY_K more (32 in all, as before) at ONE deeper loop count from
71
+ # RL_RETRY_LOOPS, assigned at random, so proofs that only deeper sampling finds enter the
72
+ # training set and no deep sample is spent on a target one loop already solves. The random
73
+ # assignment makes the retry calls a controlled comparison of loop counts on resistant targets.
74
+ # Rounds repeat until another one would not fit in the budget.
75
+ RL_FIRST_LOOPS, RL_K = 1, 8
76
+ RL_RETRY_LOOPS, RL_RETRY_K = (2, 4, 8), 24
77
+ # The fine-tune sees FT_RECORDS records per round (the old 600 steps x 128 records) in the same
78
+ # token-budget batches as pretraining (TOK_BUDGET below) instead of 128-record batches padded to
79
+ # their longest proof, which with 7-14 line RL proofs in nearly every batch was ~3x the content.
80
+ FT_RECORDS, FT_LR, FT_WARMUP = 600 * 128, 3e-4, 50
81
+ RETAIN, MAX_PER_THM, RL_WEIGHT = 20000, 4, 4
82
+ # MLP width held at 151's 1280: the fork's Block hardcodes 4 * D (1536 at D 384 would pass the
83
+ # 10M cap), so PeriLNBlock swaps fc1/fc2. 1280 != 3 * D keeps muon_step's QKV-by-shape test valid.
84
+ D_FF = 1280
85
+ BS, LR, MIN_LR_FRAC, WD = 128, 1e-3, 0.3, 0.1
86
+ MUON_LR, MUON_MOMENTUM = 0.01, 0.95
87
+ # Polar Express (Amsel et al. 2025) per-iteration quintic coefficients, as in modded-nanogpt
88
+ # (num_iters=5, safety factor 2e-2): singular values land in ~[0.86, 1.14] instead of the
89
+ # fixed quintic's ~[0.68, 1.13], at the same five iterations.
90
+ NS_COEFFS = (
91
+ (8.156554524902461, -22.48329292557795, 15.878769915207462),
92
+ (4.042929935166739, -2.808917465908714, 0.5000178451051316),
93
+ (3.8916678022926607, -2.772484153217685, 0.5060648178503393),
94
+ (3.285753657755655, -2.3681294933425376, 0.46449024233003106),
95
+ (2.3465413258596377, -1.7097828382687081, 0.42323551169305323),
96
+ )
97
+ # Ten times the AdamW group's per-step multiplicative decay (lr * wd) on the Muon matrices.
98
+ MUON_WD = 10 * WD * LR / MUON_LR
99
+ TOK_MODE = "lean_seq" # required by the harness (Lean protocol)
100
+ # Hybrid heads: half the heads get ALiBi (Press et al. 2021) with the standard slope set for
101
+ # H/2 heads, 2^(-8 i / (H/2)) = 1/4 .. 1/256 with 8 heads; the other half get slope 0, i.e. no
102
+ # positional signal at all (NoPE), so they can look citations up by content at any distance.
103
+ N_ALIBI = N_HEAD // 2
104
+ ALIBI_SLOPES = tuple(2.0 ** (-8.0 * (i + 1) / N_ALIBI) for i in range(N_ALIBI)) + (0.0,) * (
105
+ N_HEAD - N_ALIBI
106
+ )
107
+ VAL_N, CURVE_EVERY = 1024, 15 # records held out for val_loss; seconds between curve points
108
+ # Token-budget batches (fairseq max-tokens): each step takes a run of length-sorted records with
109
+ # count * padded_length <= TOK_BUDGET, so a batch of short proofs has many records and a batch of
110
+ # long ones few, and the padded token count per step is ~constant. 128 x 144 matches the real
111
+ # (unpadded) tokens of the incumbent's 128-record batches (mean record ~135 tokens), whose
112
+ # padding to the batch max (~400) made every step ~3x larger than its content.
113
+ TOK_BUDGET = 128 * 144
114
+ # Multi-token prediction (DeepSeek-V3 2024, Gloeckle et al. 2024), pretraining only: one extra
115
+ # Peri-LN block reads [trunk state at t; embedding of token t+1] and predicts token t+2 through
116
+ # the model's own ln_f/head. Its loss is added at this weight (DeepSeek-V3's early-training value).
117
+ MTP_WEIGHT = 0.3
118
+
119
+
120
+ # Muon update adapted from KellerJordan/Muon, commit
121
+ # f98f1cacc0263b04290753e32be8d498c1efc806 (muon.py).
122
+ # MIT License
123
+ # Copyright (c) 2024 Keller Jordan
124
+ #
125
+ # Permission is hereby granted, free of charge, to any person obtaining a copy
126
+ # of this software and associated documentation files (the "Software"), to deal
127
+ # in the Software without restriction, including without limitation the rights
128
+ # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
129
+ # copies of the Software, and to permit persons to whom the Software is
130
+ # furnished to do so, subject to the following conditions:
131
+ # The above copyright notice and this permission notice shall be included in all
132
+ # copies or substantial portions of the Software.
133
+ # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
134
+ # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
135
+ # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
136
+ # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
137
+ # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
138
+ # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
139
+ # SOFTWARE.
140
+ @torch.no_grad()
141
+ def muon_step(params, buffers, lr):
142
+ """Nesterov momentum plus five bf16 Polar Express steps, on hidden matrices only."""
143
+ for p, momentum in zip(params, buffers, strict=True):
144
+ if p.grad is None:
145
+ continue
146
+ momentum.lerp_(p.grad, 1 - MUON_MOMENTUM)
147
+ update = p.grad.lerp(momentum, MUON_MOMENTUM).bfloat16()
148
+ # Only QKV has shape (3D, D) in this GPT; orthogonalize its roles separately.
149
+ if update.size(0) == 3 * update.size(1):
150
+ update = update.reshape(3, update.size(1), update.size(1))
151
+ transpose = update.size(-2) > update.size(-1)
152
+ if transpose:
153
+ update = update.mT
154
+ update = update / (update.norm(dim=(-2, -1), keepdim=True) * (1 + 2e-2) + 1e-6)
155
+ for a, b, c in NS_COEFFS:
156
+ gram = update @ update.mT
157
+ polynomial = b * gram + c * gram @ gram
158
+ update = a * update + polynomial @ update
159
+ if transpose:
160
+ update = update.mT
161
+ p.mul_(1 - lr * MUON_WD)
162
+ # lr may be a 0-d tensor (CUDA-graphed step); scale in p's dtype like add_(alpha=-lr)
163
+ p.sub_(lr * update.reshape_as(p).to(p.dtype))
164
+
165
+
166
+ class PeriLNBlock(Block):
167
+ """P22 Peri-LN (Kim et al. 2025): the fork's pre-LN block plus a LayerNorm on each branch
168
+ output, so every residual write has unit scale times a learned gain. Attention has no RoPE:
169
+ `mask` is the additive ALiBi bias (B or 1, H, T, S) that LoopedGPT builds once per forward
170
+ and cos/sin are unused."""
171
+
172
+ def __init__(self, d, h):
173
+ super().__init__(d, h)
174
+ self.fc1, self.fc2 = torch.nn.Linear(d, D_FF), torch.nn.Linear(D_FF, d)
175
+ self.ln_attn_out, self.ln_mlp_out = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
176
+
177
+ def attn(self, x, cos, sin, mask, cache):
178
+ B, T, D = x.shape
179
+ q, k, v = self.qkv(x).view(B, T, 3, self.h, D // self.h).permute(2, 0, 3, 1, 4)
180
+ if cache is not None: # the fork's preallocated KV cache, unchanged
181
+ if "k" not in cache:
182
+ cache["k"] = k.new_empty(B, self.h, cache["max"], D // self.h)
183
+ cache["v"] = v.new_empty(B, self.h, cache["max"], D // self.h)
184
+ cache["n"] = 0
185
+ n0, n1 = cache["n"], cache["n"] + T
186
+ cache["k"][:, :, n0:n1] = k
187
+ cache["v"][:, :, n0:n1] = v
188
+ cache["n"] = n1
189
+ k, v = cache["k"][:, :, :n1], cache["v"][:, :, :n1]
190
+ y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask.to(q.dtype))
191
+ return self.proj(y.transpose(1, 2).reshape(B, T, D))
192
+
193
+ def forward(self, x, cos, sin, mask=None, cache=None):
194
+ x = x + self.ln_attn_out(self.attn(self.ln1(x), cos, sin, mask, cache))
195
+ return x + self.ln_mlp_out(self.fc2(torch.nn.functional.gelu(self.fc1(self.ln2(x)))))
196
+
197
+
198
+ class LoopedGPT(GPT):
199
+ """Looped trunk (Universal Transformer, Dehghani et al. 2018; sandwich layout and input
200
+ injection as in Geiping et al. 2025): blocks[0] is the entry block, blocks[1:-1] the shared
201
+ core applied n times, blocks[-1] the exit block. Every core pass starts from
202
+ ln_loop(state) + e, with e the entry block's output, so each pass sees an input of the same
203
+ scale and the prompt is re-injected instead of being buried under accumulated writes.
204
+
205
+ P34: the fork's GPT with ALiBi (Press et al. 2021) instead of RoPE. Head h's attention
206
+ logits get -slope_h * (query position - key position), slopes 2^-1 .. 2^-8 for 8 heads, so
207
+ no rotation of q/k anywhere. Positions come from `pos` (sample.generate's left-padded
208
+ batches) or arange; key positions are kept in caches[0] so cached decoding sees the same
209
+ bias as the full forward. Causality/padding come from `mask` when given, else from rel < 0."""
210
+
211
+ def __init__(self, vocab, n_layer, d, n_head, seed, max_len=1024):
212
+ super().__init__(vocab, n_layer, d, n_head, max_len)
213
+ slopes = torch.tensor(ALIBI_SLOPES).view(1, n_head, 1, 1)
214
+ self.register_buffer("slopes", slopes, persistent=False)
215
+ self.ln_loop = torch.nn.LayerNorm(d)
216
+ self.n_loops, self.loop_rng = DEFAULT_LOOPS, random.Random(seed)
217
+
218
+ def forward(self, idx, pos=None, mask=None, caches=None):
219
+ return self.head(self.ln_f(self.features(idx, pos, mask, caches)[0]))
220
+
221
+ def features(self, idx, pos=None, mask=None, caches=None):
222
+ """The trunk: (final hidden state before ln_f, the ALiBi bias used)."""
223
+ B, T = idx.shape
224
+ if pos is None: # training/val: one (1, H, T, T) bias broadcast over the batch
225
+ pos = torch.arange(T, device=idx.device)[None]
226
+ key_pos = pos
227
+ if caches is not None: # incremental decoding: keys are every cached token
228
+ c = caches[0]
229
+ if "pos" not in c:
230
+ c["pos"], c["n_pos"] = pos.new_empty(pos.size(0), c["max"]), 0
231
+ n0, n1 = c["n_pos"], c["n_pos"] + T
232
+ c["pos"][:, n0:n1] = pos
233
+ c["n_pos"] = n1
234
+ key_pos = c["pos"][:, :n1]
235
+ rel = pos[:, None, :, None] - key_pos[:, None, None, :] # (B or 1, 1, T, S)
236
+ bias = -self.slopes * rel.to(self.slopes.dtype)
237
+ bias = bias.masked_fill(rel < 0 if mask is None else ~mask, float("-inf"))
238
+ n = self.loop_rng.choice(TRAIN_LOOPS) if self.training else self.n_loops
239
+ entry, *core, exit_block = self.blocks
240
+ x = e = entry(self.emb(idx), None, None, bias, None if caches is None else caches[0])
241
+ for loop in range(n):
242
+ x = self.ln_loop(x) + e
243
+ for j, b in enumerate(core, 1):
244
+ # sample.generate makes one cache dict per unique block; a shared block's keys
245
+ # and values differ per pass, so each pass gets its own sub-cache inside it.
246
+ c = caches and caches[j].setdefault(loop, {"max": caches[j]["max"]})
247
+ x = b(x, None, None, bias, c)
248
+ x = exit_block(x, None, None, bias, None if caches is None else caches[-1])
249
+ return x, bias
250
+
251
+
252
+ class MTPHead(torch.nn.Module):
253
+ """The MTP module. A separate nn.Module, never registered on the model, so the harness's
254
+ parameter count, RL AdamW and sampler never see it; it is dropped when pretrain() returns."""
255
+
256
+ def __init__(self, d, h):
257
+ super().__init__()
258
+ self.ln_h, self.ln_e = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
259
+ self.proj = torch.nn.Linear(2 * d, d, bias=False)
260
+ self.block = PeriLNBlock(d, h)
261
+
262
+
263
+ def mtp_loss(model, mtp, x, m):
264
+ """The fork's loss_on (next-token CE on proof tokens) plus MTP_WEIGHT times the CE of the
265
+ MTP module's token-(t+2) prediction at the same proof positions. Returns (total, main)."""
266
+ h, bias = model.features(x[:, :-1])
267
+ logits = model.head(model.ln_f(h))
268
+ V = logits.size(-1)
269
+ ce = torch.nn.functional.cross_entropy(
270
+ logits.reshape(-1, V).float(), x[:, 1:].reshape(-1), reduction="none"
271
+ )
272
+ main = (ce * m[:, 1:].reshape(-1)).sum() / m[:, 1:].sum()
273
+ # Position t (0..T-3) gets h_t and emb(x_{t+1}); causal attention over these keeps every
274
+ # prediction of x_{t+2} a function of x_{<=t+1}.
275
+ z = mtp.proj(torch.cat((mtp.ln_h(h[:, :-1]), mtp.ln_e(model.emb(x[:, 1:-1]))), -1))
276
+ z = mtp.block(z, None, None, bias[..., :-1, :-1], None)
277
+ ce2 = torch.nn.functional.cross_entropy(
278
+ model.head(model.ln_f(z)).reshape(-1, V).float(), x[:, 2:].reshape(-1), reduction="none"
279
+ )
280
+ aux = (ce2 * m[:, 2:].reshape(-1)).sum() / m[:, 2:].sum().clamp(min=1)
281
+ return main + MTP_WEIGHT * aux, main, aux
282
+
283
+
284
+ def val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast):
285
+ """Mean loss over the first VAL_N records (never trained on), no shift augmentation."""
286
+ model.eval()
287
+ tot = 0.0
288
+ with torch.no_grad(), autocast:
289
+ for i in range(0, VAL_N, BS):
290
+ idxs = torch.arange(i, min(i + BS, VAL_N), device=ids.device)
291
+ x, m = ids[idxs], mask[idxs]
292
+ T = int(lens[idxs].max())
293
+ tot += loss_on(model, x[:, :T], m[:, :T]).item() * len(idxs)
294
+ model.train()
295
+ return tot / VAL_N
296
+
297
+
298
+ def to_gpu(data, tok, device):
299
+ """Pad all (prompt, proof) records once into (N, T) ids + proof mask on the device."""
300
+ N, T = len(data), max(len(p) + len(q) for p, q in data)
301
+ ids = torch.full((N, T), tok.pad, dtype=torch.long)
302
+ mask = torch.zeros((N, T), dtype=torch.bool)
303
+ ref_max = torch.zeros(
304
+ N, dtype=torch.long
305
+ ) # largest N<k> index used in the proof, for the shift augmentation
306
+ for i, (p, q) in enumerate(data):
307
+ ids[i, : len(p) + len(q)] = torch.tensor(p + q)
308
+ mask[i, len(p) : len(p) + len(q)] = True
309
+ ref_max[i] = max((x - tok.ref0 + 1 for x in q if x >= tok.ref0), default=0)
310
+ return ids.to(device), mask.to(device), ref_max.to(device), (ids != tok.pad).sum(1).to(device)
311
+
312
+
313
+ def token_batches(lens, n_train, gen, rng, first=VAL_N):
314
+ """One epoch of token-budget batches: a random permutation of the n_train records from
315
+ `first` on, stably sorted by length (equal lengths stay shuffled), cut greedily into runs of
316
+ consecutive records with count * max_len <= TOK_BUDGET, and the runs' order shuffled.
317
+ Returns (order, cuts)."""
318
+ perm = torch.randperm(n_train, device=lens.device, generator=gen) + first
319
+ order = perm[torch.argsort(lens[perm], stable=True)]
320
+ sorted_lens = lens[order].cpu()
321
+ cuts, s = [], 0
322
+ while s < len(sorted_lens):
323
+ window = sorted_lens[s : s + TOK_BUDGET // int(sorted_lens[s])]
324
+ # count * max_len is non-decreasing along the window, so the fits form a prefix.
325
+ e = s + int((torch.arange(1, len(window) + 1) * window <= TOK_BUDGET).sum())
326
+ cuts.append((s, e, int(sorted_lens[e - 1]))) # sorted: last is the batch's max length
327
+ s = e
328
+ rng.shuffle(cuts)
329
+ return order, cuts
330
+
331
+
332
+ def gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, max_len=None):
333
+ """The fork's batch() + tok.shift_abs, vectorised: a random per-row offset is added to
334
+ every ref token."""
335
+ x, m = ids[idxs], mask[idxs]
336
+ # Known on the host for training batches: passing it avoids a GPU sync every step.
337
+ T = int(lens[idxs].max()) if max_len is None else max_len
338
+ x, m = x[:, :T].clone(), m[:, :T]
339
+ if tok.mode in ("abs", "lean_seq"): # lean_seq: names n1.. get the same start offset
340
+ room = MAXN - ref_max[idxs]
341
+ shift = (
342
+ (torch.rand(len(idxs), device=x.device, generator=gen) * (room + 1))
343
+ .long()
344
+ .clamp(max=room)
345
+ )
346
+ is_ref = (x >= tok.ref0) & m
347
+ x = torch.where(is_ref, x + shift[:, None], x)
348
+ return x, m
349
+
350
+
351
+ def pretrain(train_recs, budget_secs, device, seed, log):
352
+ t0 = time.time()
353
+ tok = make_tokenizer(TOK_MODE)
354
+ data = [(tok.encode_prompt(r["prompt"]), tok.encode_proof(r["proof"])) for r in train_recs]
355
+ ids, mask, ref_max, lens = to_gpu(data, tok, device)
356
+ gen = torch.Generator(device=device)
357
+ gen.manual_seed(seed)
358
+ model = LoopedGPT(tok.vocab_size, N_LAYER, D, N_HEAD, seed)
359
+ model.blocks = torch.nn.ModuleList(PeriLNBlock(D, N_HEAD) for _ in range(N_LAYER))
360
+ model.blocks.apply(model._init) # fork init: zero biases
361
+ model = model.to(device)
362
+ mtp = MTPHead(D, N_HEAD)
363
+ mtp.apply(model._init)
364
+ mtp = mtp.to(device)
365
+ # Compile the repeated blocks once each (torch >= 2.9; dynamic shapes, ~14 s): ~14% faster
366
+ # steps. The coalescing tiling analysis crashes on dynamic shapes, so it is off. Undone at
367
+ # handoff so the harness's RL fine-tune and KV-cache sampling run eager.
368
+ torch._inductor.config.triton.coalesce_tiling_analysis = False
369
+ muon_lr, muon_graph = torch.zeros((), device=device), None
370
+ compiled = list(model.blocks) + [mtp.block]
371
+ for b in compiled if device == "cuda" else []: # CPU smoke runs stay eager
372
+ b.compile(dynamic=True)
373
+ for p in list(model.blocks.parameters()) + list(mtp.parameters()):
374
+ if p.ndim == 2:
375
+ # Orthogonal matrices have element second moment gain^2 / max(rows, cols).
376
+ torch.nn.init.orthogonal_(p, gain=0.02 * math.sqrt(max(p.shape)))
377
+ log(
378
+ f"pretrain: {model.n_params():,} params, {len(data):,} records on {device} "
379
+ f"after {time.time() - t0:.0f}s, budget {budget_secs}s"
380
+ )
381
+ # The output readout joins the hidden matrices under Muon; embeddings stay on AdamW.
382
+ trainable = list(model.parameters()) + list(mtp.parameters())
383
+ matrices = [p for p in model.blocks.parameters() if p.ndim == 2] + [model.head.weight]
384
+ matrices += [p for p in mtp.parameters() if p.ndim == 2]
385
+ matrix_ids = {id(p) for p in matrices}
386
+ auxiliary = [p for p in trainable if id(p) not in matrix_ids]
387
+ momentum_buffers = [torch.zeros_like(p) for p in matrices]
388
+ opt = torch.optim.AdamW(
389
+ auxiliary, lr=LR, weight_decay=WD, betas=(0.9, 0.95), fused=device == "cuda"
390
+ )
391
+ autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
392
+ model.train()
393
+ step, last_loss, records_seen = 0, float("nan"), 0
394
+ curve, recent, recent_aux, next_point = [], [], [], CURVE_EVERY
395
+ n_train = len(data) - VAL_N # records [0, VAL_N) are the val slice
396
+ rng, cuts, batches_per_epoch = random.Random(seed), [], None
397
+ order = torch.empty(0, dtype=torch.long, device=device)
398
+ while True:
399
+ frac = (
400
+ time.time() - t0
401
+ ) / budget_secs # cosine over wall-clock, so the cap is the schedule
402
+ if frac >= 1.0:
403
+ break
404
+ step += 1
405
+ # No warmup (as in modded-nanogpt): Muon's step is normalized and Adam's is ~lr per
406
+ # element from step 1, and Peri-LN normalizes every branch write.
407
+ lr = LR * MIN_LR_FRAC + 0.5 * LR * (1 - MIN_LR_FRAC) * (1 + math.cos(math.pi * frac))
408
+ for g in opt.param_groups:
409
+ g["lr"] = lr
410
+ if not cuts: # new epoch
411
+ order, cuts = token_batches(lens, n_train, gen, rng)
412
+ batches_per_epoch = len(cuts)
413
+ s, e, T = cuts.pop()
414
+ idxs = order[s:e]
415
+ records_seen += e - s
416
+ x, m = gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, T)
417
+ with autocast:
418
+ loss, main, aux = mtp_loss(model, mtp, x, m)
419
+ model.zero_grad(set_to_none=False) # grads keep fixed addresses for the Muon CUDA graph
420
+ mtp.zero_grad(set_to_none=False)
421
+ loss.backward()
422
+ torch.nn.utils.clip_grad_norm_(trainable, 1.0)
423
+ opt.step()
424
+ muon_lr.fill_(MUON_LR * lr / LR)
425
+ # Capture once; later steps replay one graph instead of ~400 kernel launches. CPU smoke
426
+ # runs call the step directly.
427
+ if device != "cuda":
428
+ muon_step(matrices, momentum_buffers, muon_lr)
429
+ elif muon_graph is None:
430
+ state = matrices + momentum_buffers
431
+ saved = [t.clone() for t in state]
432
+ side = torch.cuda.Stream()
433
+ side.wait_stream(torch.cuda.current_stream())
434
+ with torch.cuda.stream(side):
435
+ muon_step(matrices, momentum_buffers, muon_lr) # warm-up (lazy cuBLAS init)
436
+ torch.cuda.current_stream().wait_stream(side)
437
+ muon_graph = torch.cuda.CUDAGraph()
438
+ with torch.cuda.graph(muon_graph):
439
+ muon_step(matrices, momentum_buffers, muon_lr)
440
+ with torch.no_grad():
441
+ for t, v in zip(state, saved, strict=True):
442
+ t.copy_(v) # undo the warm-up and capture updates
443
+ muon_graph.replay()
444
+ else:
445
+ muon_graph.replay()
446
+ if step % 20 == 0:
447
+ last_loss = main.item() # the next-token CE, comparable with earlier runs
448
+ recent.append(last_loss)
449
+ recent_aux.append(aux.item())
450
+ if (
451
+ time.time() - t0 >= next_point
452
+ ): # curve point: mean recent train loss + val loss on the held-out slice
453
+ curve.append(
454
+ {
455
+ "secs": round(time.time() - t0),
456
+ "step": step,
457
+ "train_loss": sum(recent) / len(recent) if recent else last_loss,
458
+ "val_loss": val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast),
459
+ "mtp_loss": sum(recent_aux) / len(recent_aux) if recent_aux else None,
460
+ }
461
+ )
462
+ log(
463
+ f"pretrain step {step} train {curve[-1]['train_loss']:.4f} "
464
+ f"val {curve[-1]['val_loss']:.4f} lr {lr:.2e} {curve[-1]['secs']}s"
465
+ )
466
+ recent, recent_aux, next_point = [], [], next_point + CURVE_EVERY
467
+ val_by_loops = {}
468
+ for n in VAL_LOOPS if step else (): # no steps: a checkpoint rebuild, not a training run
469
+ model.n_loops = n
470
+ val_by_loops[n] = val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast)
471
+ model.n_loops = DEFAULT_LOOPS
472
+ log(f"pretrain val loss by loops: {val_by_loops}")
473
+ for b in compiled:
474
+ b._compiled_call_impl = None
475
+ model.eval()
476
+ return (
477
+ model,
478
+ tok,
479
+ {
480
+ "steps": step,
481
+ "records_seen": records_seen,
482
+ "epochs": records_seen / n_train,
483
+ "batches_per_epoch": batches_per_epoch,
484
+ "tok_budget": TOK_BUDGET,
485
+ "last_loss": last_loss,
486
+ "train_loops": TRAIN_LOOPS,
487
+ "default_loops": DEFAULT_LOOPS,
488
+ "val_by_loops": val_by_loops,
489
+ "curve": curve,
490
+ },
491
+ )
492
+
493
+
494
+ def finetune(model, tok, data, rng, device):
495
+ """The old harness's RL fine-tune: AdamW, linear warmup then cosine to lr/10, over FT_RECORDS
496
+ records in token-budget batches (to_gpu / token_batches / gpu_batch, as pretraining). The
497
+ model is in train mode, so every batch draws its own loop count from TRAIN_LOOPS. Returns
498
+ (last loss, steps)."""
499
+ ids, mask, ref_max, lens = to_gpu(data, tok, device)
500
+ gen = torch.Generator(device=device)
501
+ gen.manual_seed(rng.getrandbits(31))
502
+ plan, seen = [], 0 # (order, start, end, padded length) per step, epoch after epoch
503
+ while seen < FT_RECORDS:
504
+ order, cuts = token_batches(lens, len(data), gen, rng, first=0)
505
+ for s, e, T in cuts:
506
+ if seen >= FT_RECORDS:
507
+ break
508
+ plan.append((order, s, e, T))
509
+ seen += e - s
510
+ opt = torch.optim.AdamW(model.parameters(), lr=FT_LR, weight_decay=0.1, betas=(0.9, 0.95))
511
+ autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
512
+ model.train()
513
+ for step, (order, s, e, T) in enumerate(plan, 1):
514
+ lr = (
515
+ FT_LR * step / FT_WARMUP
516
+ if step < FT_WARMUP
517
+ else FT_LR / 10 + 0.45 * FT_LR * (1 + math.cos(math.pi * step / len(plan)))
518
+ )
519
+ for g in opt.param_groups:
520
+ g["lr"] = lr
521
+ x, m = gpu_batch(ids, mask, ref_max, lens, order[s:e], tok, gen, T)
522
+ with autocast:
523
+ loss = loss_on(model, x, m)
524
+ opt.zero_grad(set_to_none=True)
525
+ loss.backward()
526
+ torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
527
+ opt.step()
528
+ model.eval()
529
+ return loss.item(), len(plan)
530
+
531
+
532
+ def rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log):
533
+ rng = random.Random(seed)
534
+ found = {t["name"]: {} for t in targets} # name -> {normalized proof: proof text}
535
+ rounds, round_secs, retries, ft_steps = 0, 0.0, [], []
536
+ # Start a round only if one as long as the last (plus 15%) still fits in the budget.
537
+ while oracle.left() > 1.15 * round_secs:
538
+ t0 = oracle.used()
539
+ model.n_loops = RL_FIRST_LOOPS
540
+ rows = oracle.sample(model, targets, RL_K)
541
+ failed = [t for t, row in zip(targets, rows, strict=True) if not row["proofs"]]
542
+ rng.shuffle(failed)
543
+ for i, n in enumerate(RL_RETRY_LOOPS):
544
+ part = failed[i :: len(RL_RETRY_LOOPS)]
545
+ if not part:
546
+ continue
547
+ model.n_loops = n
548
+ part_rows = oracle.sample(model, part, RL_RETRY_K)
549
+ cracked = sum(bool(row["proofs"]) for row in part_rows)
550
+ retries.append(
551
+ {"round": rounds + 1, "loops": n, "targets": len(part), "cracked": cracked}
552
+ )
553
+ rows += part_rows
554
+ for row in rows:
555
+ for proof in row["proofs"]:
556
+ found[row["name"]].setdefault(norm(proof), proof)
557
+ mix = []
558
+ for t in targets:
559
+ proofs = list(found[t["name"]].values())
560
+ rng.shuffle(proofs)
561
+ mix += [(t["prompt"], proof) for proof in proofs[:MAX_PER_THM]] * RL_WEIGHT
562
+ n_rl = len(mix)
563
+ if not n_rl:
564
+ break
565
+ mix += [(r["prompt"], r["proof"]) for r in rng.sample(train_recs, RETAIN)]
566
+ data = [(tok.encode_prompt(p), tok.encode_proof(q)) for p, q in mix]
567
+ loss, steps = finetune(model, tok, data, rng, device)
568
+ ft_steps.append(steps)
569
+ rounds += 1
570
+ round_secs = oracle.used() - t0
571
+ log(
572
+ f"rl round {rounds}: {n_rl} RL records, fine-tune loss {loss:.4f} in {steps} steps, "
573
+ f"round {round_secs:.0f}s, {oracle.left():.0f}s left"
574
+ )
575
+ model.n_loops = DEFAULT_LOOPS
576
+ return {
577
+ "rounds": rounds,
578
+ "first_loops": RL_FIRST_LOOPS,
579
+ "k_first": RL_K,
580
+ "retry_loops": RL_RETRY_LOOPS,
581
+ "k_retry": RL_RETRY_K,
582
+ "retries": retries,
583
+ "ft_records": FT_RECORDS,
584
+ "ft_steps": ft_steps,
585
+ }
runs/507-fable_ft-token-batches_s1/source/remote.sh ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Runs on the pod. Checks out Dan's fork at the pinned commit (frozen data + verifier + Lean
3
+ # tokenizer/gate), installs Lean, runs the harness.
4
+ # Usage: bash remote.sh <out-dir> [harness.py args...]
5
+ set -euo pipefail
6
+ FORK_COMMIT=51604b3feaa5e66821d1faba22238fe6d96d46d0 # dan-pandori/nd-takehome, branch dan_lean_seed2 (lean_seq with the BOTE fix)
7
+ export ND_FORK=/workspace/nd-takehome
8
+ OUT=$1; shift
9
+
10
+ if [ ! -d "$ND_FORK/.git" ]; then
11
+ git clone -q https://github.com/dan-pandori/nd-takehome.git "$ND_FORK"
12
+ fi
13
+ git -C "$ND_FORK" cat-file -e "$FORK_COMMIT" 2>/dev/null || git -C "$ND_FORK" fetch -q origin
14
+ git -C "$ND_FORK" checkout -q "$FORK_COMMIT"
15
+ [ -f "$ND_FORK/data/train.jsonl" ] || gunzip -k "$ND_FORK/data/train.jsonl.gz"
16
+ # Lean 4 core checks every sampled proof (fork lean_gate.py, called from sample.generate): the
17
+ # reward is Lean AND nd_verify. Installed before harness.py starts, so it is not charged to TOTAL_SECS.
18
+ [ -x ~/.elan/bin/lean ] || curl -sSf https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh \
19
+ | sh -s -- -y --default-toolchain leanprover/lean4:v4.34.0 > /dev/null
20
+ # One Lean process per CPU: on these pods (nproc 96) 12 workers took 146 s per RL-round gate
21
+ # pass (099) against 67-84 s with 96 workers (098).
22
+ export LEAN_GATE_WORKERS=$(nproc) LEAN_GATE_LOG="$OUT/lean_gate.jsonl"
23
+ nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
24
+ echo "cpus $(nproc), $(~/.elan/bin/lean --version)"
25
+ # Newer RunPod images put torch under `python` and leave `python3` as bare system Python.
26
+ PY=$(for p in python python3; do "$p" -c "import torch" 2>/dev/null && echo "$p" && break; done)
27
+ echo "python $($PY --version 2>&1), torch $($PY -c 'import torch; print(torch.__version__)')"
28
+ cd /workspace/autoresearch
29
+ $PY harness.py --out "$OUT" "$@" 2>&1 | tee "$OUT.log"
runs/507-fable_ft-token-batches_s1/summary.json ADDED
@@ -0,0 +1,1913 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "ok",
3
+ "metric": 765,
4
+ "metric_diag": 743,
5
+ "metric_rl": 1264,
6
+ "harness_sha": "2eec1b61666a",
7
+ "pretrain_sha": "fe991500c215",
8
+ "device": "cuda",
9
+ "gpu": "NVIDIA A40",
10
+ "torch": "2.9.1+cu129",
11
+ "lean_gate": {
12
+ "samples": 412872,
13
+ "parse_fail": 106357,
14
+ "distinct_checked": 277366,
15
+ "both_ok": 131982,
16
+ "nd_ok_lean_rej": 0,
17
+ "nd_rej_lean_ok": 2,
18
+ "lean_wall_s": 789.6761934757233
19
+ },
20
+ "total_secs": 3775.402419567108,
21
+ "rules": {
22
+ "TOTAL_SECS": 5400,
23
+ "PRETRAIN_SECS": 300,
24
+ "RL_SECS": 1800,
25
+ "PARAM_CAP": 10000000,
26
+ "N_TARGETS": 1500,
27
+ "EVAL_K": 64,
28
+ "KNOB": "n_loops",
29
+ "EVAL_LEVEL": 8,
30
+ "DIAG_LEVEL": 2,
31
+ "TOK_MODE": "lean_seq"
32
+ },
33
+ "pretrain": {
34
+ "steps": 3443,
35
+ "records_seen": 468050,
36
+ "epochs": 3.039956873595469,
37
+ "batches_per_epoch": 1132,
38
+ "tok_budget": 18432,
39
+ "last_loss": 0.04709014669060707,
40
+ "train_loops": [
41
+ 1,
42
+ 2,
43
+ 3,
44
+ 4,
45
+ 5,
46
+ 6,
47
+ 7,
48
+ 8
49
+ ],
50
+ "default_loops": 2,
51
+ "val_by_loops": {
52
+ "1": 0.08804549090564251,
53
+ "2": 0.08679556474089622,
54
+ "4": 0.08629628550261259,
55
+ "8": 0.08621826767921448
56
+ },
57
+ "curve": [
58
+ {
59
+ "secs": 30,
60
+ "step": 1,
61
+ "train_loss": NaN,
62
+ "val_loss": 3.9690618216991425,
63
+ "mtp_loss": null
64
+ },
65
+ {
66
+ "secs": 32,
67
+ "step": 2,
68
+ "train_loss": NaN,
69
+ "val_loss": 3.4871207773685455,
70
+ "mtp_loss": null
71
+ },
72
+ {
73
+ "secs": 45,
74
+ "step": 166,
75
+ "train_loss": 0.5545947533100843,
76
+ "val_loss": 0.12264137528836727,
77
+ "mtp_loss": 0.5321323983371258
78
+ },
79
+ {
80
+ "secs": 60,
81
+ "step": 367,
82
+ "train_loss": 0.12384326383471489,
83
+ "val_loss": 0.09338787849992514,
84
+ "mtp_loss": 0.11772393770515918
85
+ },
86
+ {
87
+ "secs": 75,
88
+ "step": 556,
89
+ "train_loss": 0.09290478461318546,
90
+ "val_loss": 0.0917009599506855,
91
+ "mtp_loss": 0.09065078447262447
92
+ },
93
+ {
94
+ "secs": 90,
95
+ "step": 755,
96
+ "train_loss": 0.08195694945752621,
97
+ "val_loss": 0.09133180417120457,
98
+ "mtp_loss": 0.08215314075350762
99
+ },
100
+ {
101
+ "secs": 105,
102
+ "step": 945,
103
+ "train_loss": 0.07310649119317532,
104
+ "val_loss": 0.09970938041806221,
105
+ "mtp_loss": 0.07427530996501445
106
+ },
107
+ {
108
+ "secs": 120,
109
+ "step": 1140,
110
+ "train_loss": 0.0763938345015049,
111
+ "val_loss": 0.078827403485775,
112
+ "mtp_loss": 0.07754262536764145
113
+ },
114
+ {
115
+ "secs": 135,
116
+ "step": 1332,
117
+ "train_loss": 0.06993935174412197,
118
+ "val_loss": 0.08748125843703747,
119
+ "mtp_loss": 0.07078416728311115
120
+ },
121
+ {
122
+ "secs": 150,
123
+ "step": 1527,
124
+ "train_loss": 0.07024346329271794,
125
+ "val_loss": 0.08945621736347675,
126
+ "mtp_loss": 0.07223650366067887
127
+ },
128
+ {
129
+ "secs": 165,
130
+ "step": 1724,
131
+ "train_loss": 0.057505343481898306,
132
+ "val_loss": 0.08637734223157167,
133
+ "mtp_loss": 0.05790973044931889
134
+ },
135
+ {
136
+ "secs": 180,
137
+ "step": 1917,
138
+ "train_loss": 0.07018767380052143,
139
+ "val_loss": 0.08480384293943644,
140
+ "mtp_loss": 0.07093631103634834
141
+ },
142
+ {
143
+ "secs": 195,
144
+ "step": 2100,
145
+ "train_loss": 0.06114649586379528,
146
+ "val_loss": 0.08211605437099934,
147
+ "mtp_loss": 0.06202838271856308
148
+ },
149
+ {
150
+ "secs": 210,
151
+ "step": 2293,
152
+ "train_loss": 0.05954717223842939,
153
+ "val_loss": 0.08690226636826992,
154
+ "mtp_loss": 0.05964854649371571
155
+ },
156
+ {
157
+ "secs": 225,
158
+ "step": 2488,
159
+ "train_loss": 0.06698438785970211,
160
+ "val_loss": 0.08723815716803074,
161
+ "mtp_loss": 0.06717289015650749
162
+ },
163
+ {
164
+ "secs": 240,
165
+ "step": 2685,
166
+ "train_loss": 0.052734899520874026,
167
+ "val_loss": 0.08414971362799406,
168
+ "mtp_loss": 0.05282359905540943
169
+ },
170
+ {
171
+ "secs": 255,
172
+ "step": 2865,
173
+ "train_loss": 0.04858669307496813,
174
+ "val_loss": 0.08921525999903679,
175
+ "mtp_loss": 0.04859017415179147
176
+ },
177
+ {
178
+ "secs": 270,
179
+ "step": 3059,
180
+ "train_loss": 0.06436175232132275,
181
+ "val_loss": 0.08473679423332214,
182
+ "mtp_loss": 0.06501536940534909
183
+ },
184
+ {
185
+ "secs": 285,
186
+ "step": 3251,
187
+ "train_loss": 0.05306601077318192,
188
+ "val_loss": 0.08327818848192692,
189
+ "mtp_loss": 0.052501676976680754
190
+ },
191
+ {
192
+ "secs": 300,
193
+ "step": 3443,
194
+ "train_loss": 0.055818074569106105,
195
+ "val_loss": 0.08679556474089622,
196
+ "mtp_loss": 0.056372201442718504
197
+ }
198
+ ],
199
+ "secs": 301.0448079109192,
200
+ "n_params": 6400256,
201
+ "heldout_greedy": 0.886
202
+ },
203
+ "rl": {
204
+ "rounds": [
205
+ {
206
+ "round": 1,
207
+ "level": 1,
208
+ "k": 8,
209
+ "n_targets": 1500,
210
+ "sample_acc": 0.006666666666666667,
211
+ "new_proofs": 36,
212
+ "targets_solved": 34,
213
+ "lstar": 7,
214
+ "ge": {
215
+ "7": 34,
216
+ "8": 1,
217
+ "9": 0,
218
+ "10": 0,
219
+ "11": 0,
220
+ "12": 0,
221
+ "13": 0,
222
+ "14": 0
223
+ },
224
+ "mix_rl_records": 36,
225
+ "sampling": {
226
+ "samples": 12000,
227
+ "secs": 43.03496265411377,
228
+ "gate_secs": 9.748952627182007,
229
+ "samples_per_sec": 278.8430443508914
230
+ },
231
+ "secs": 43.23123049736023,
232
+ "rl_secs_used": 33.48254990577698
233
+ },
234
+ {
235
+ "round": 2,
236
+ "level": 2,
237
+ "k": 24,
238
+ "n_targets": 489,
239
+ "sample_acc": 0.014570552147239263,
240
+ "new_proofs": 36,
241
+ "targets_solved": 67,
242
+ "lstar": 7,
243
+ "ge": {
244
+ "7": 67,
245
+ "8": 4,
246
+ "9": 1,
247
+ "10": 0,
248
+ "11": 0,
249
+ "12": 0,
250
+ "13": 0,
251
+ "14": 0
252
+ },
253
+ "mix_rl_records": 36,
254
+ "sampling": {
255
+ "samples": 11736,
256
+ "secs": 55.05180835723877,
257
+ "gate_secs": 10.128597974777222,
258
+ "samples_per_sec": 213.1810080396175
259
+ },
260
+ "secs": 55.26160264015198,
261
+ "rl_secs_used": 78.6154043674469
262
+ },
263
+ {
264
+ "round": 3,
265
+ "level": 4,
266
+ "k": 24,
267
+ "n_targets": 489,
268
+ "sample_acc": 0.04115541922290389,
269
+ "new_proofs": 70,
270
+ "targets_solved": 120,
271
+ "lstar": 8,
272
+ "ge": {
273
+ "7": 120,
274
+ "8": 10,
275
+ "9": 4,
276
+ "10": 0,
277
+ "11": 0,
278
+ "12": 0,
279
+ "13": 0,
280
+ "14": 0
281
+ },
282
+ "mix_rl_records": 70,
283
+ "sampling": {
284
+ "samples": 11736,
285
+ "secs": 81.60133242607117,
286
+ "gate_secs": 11.293653726577759,
287
+ "samples_per_sec": 143.82118099153016
288
+ },
289
+ "secs": 81.83586192131042,
290
+ "rl_secs_used": 149.15758728981018
291
+ },
292
+ {
293
+ "round": 4,
294
+ "level": 8,
295
+ "k": 24,
296
+ "n_targets": 488,
297
+ "sample_acc": 0.040385928961748634,
298
+ "new_proofs": 67,
299
+ "targets_solved": 171,
300
+ "lstar": 9,
301
+ "ge": {
302
+ "7": 171,
303
+ "8": 22,
304
+ "9": 8,
305
+ "10": 0,
306
+ "11": 0,
307
+ "12": 0,
308
+ "13": 0,
309
+ "14": 0
310
+ },
311
+ "mix_rl_records": 67,
312
+ "sampling": {
313
+ "samples": 11712,
314
+ "secs": 106.70303416252136,
315
+ "gate_secs": 12.363486289978027,
316
+ "samples_per_sec": 109.7625769681604
317
+ },
318
+ "secs": 106.93916368484497,
319
+ "rl_secs_used": 243.73325324058533
320
+ },
321
+ {
322
+ "round": 5,
323
+ "level": 1,
324
+ "k": 8,
325
+ "n_targets": 1500,
326
+ "sample_acc": 0.31858333333333333,
327
+ "new_proofs": 670,
328
+ "targets_solved": 673,
329
+ "lstar": 10,
330
+ "ge": {
331
+ "7": 673,
332
+ "8": 255,
333
+ "9": 70,
334
+ "10": 7,
335
+ "11": 0,
336
+ "12": 0,
337
+ "13": 0,
338
+ "14": 0
339
+ },
340
+ "mix_rl_records": 853,
341
+ "sampling": {
342
+ "samples": 12000,
343
+ "secs": 49.214065074920654,
344
+ "gate_secs": 26.633244276046753,
345
+ "samples_per_sec": 243.83273321827596
346
+ },
347
+ "secs": 100.38916301727295,
348
+ "rl_secs_used": 317.4892632961273
349
+ },
350
+ {
351
+ "round": 6,
352
+ "level": 2,
353
+ "k": 24,
354
+ "n_targets": 278,
355
+ "sample_acc": 0.045563549160671464,
356
+ "new_proofs": 56,
357
+ "targets_solved": 722,
358
+ "lstar": 10,
359
+ "ge": {
360
+ "7": 722,
361
+ "8": 292,
362
+ "9": 88,
363
+ "10": 11,
364
+ "11": 0,
365
+ "12": 0,
366
+ "13": 0,
367
+ "14": 0
368
+ },
369
+ "mix_rl_records": 58,
370
+ "sampling": {
371
+ "samples": 6672,
372
+ "secs": 32.981425523757935,
373
+ "gate_secs": 17.429762601852417,
374
+ "samples_per_sec": 202.2956829198869
375
+ },
376
+ "secs": 33.23595857620239,
377
+ "rl_secs_used": 333.2953441143036
378
+ },
379
+ {
380
+ "round": 7,
381
+ "level": 4,
382
+ "k": 24,
383
+ "n_targets": 278,
384
+ "sample_acc": 0.09097721822541967,
385
+ "new_proofs": 85,
386
+ "targets_solved": 785,
387
+ "lstar": 10,
388
+ "ge": {
389
+ "7": 785,
390
+ "8": 336,
391
+ "9": 120,
392
+ "10": 16,
393
+ "11": 0,
394
+ "12": 0,
395
+ "13": 0,
396
+ "14": 0
397
+ },
398
+ "mix_rl_records": 88,
399
+ "sampling": {
400
+ "samples": 6672,
401
+ "secs": 42.920403718948364,
402
+ "gate_secs": 18.78557538986206,
403
+ "samples_per_sec": 155.4505415114366
404
+ },
405
+ "secs": 43.19931697845459,
406
+ "rl_secs_used": 357.70909452438354
407
+ },
408
+ {
409
+ "round": 8,
410
+ "level": 8,
411
+ "k": 24,
412
+ "n_targets": 277,
413
+ "sample_acc": 0.09732250300842359,
414
+ "new_proofs": 102,
415
+ "targets_solved": 854,
416
+ "lstar": 10,
417
+ "ge": {
418
+ "7": 854,
419
+ "8": 381,
420
+ "9": 148,
421
+ "10": 22,
422
+ "11": 1,
423
+ "12": 0,
424
+ "13": 0,
425
+ "14": 0
426
+ },
427
+ "mix_rl_records": 104,
428
+ "sampling": {
429
+ "samples": 6648,
430
+ "secs": 60.06416153907776,
431
+ "gate_secs": 16.93283462524414,
432
+ "samples_per_sec": 110.68164159212994
433
+ },
434
+ "secs": 60.329643964767456,
435
+ "rl_secs_used": 401.10589385032654
436
+ },
437
+ {
438
+ "round": 9,
439
+ "level": 1,
440
+ "k": 8,
441
+ "n_targets": 1500,
442
+ "sample_acc": 0.46425,
443
+ "new_proofs": 293,
444
+ "targets_solved": 923,
445
+ "lstar": 10,
446
+ "ge": {
447
+ "7": 923,
448
+ "8": 429,
449
+ "9": 186,
450
+ "10": 36,
451
+ "11": 1,
452
+ "12": 0,
453
+ "13": 0,
454
+ "14": 0
455
+ },
456
+ "mix_rl_records": 1254,
457
+ "sampling": {
458
+ "samples": 12000,
459
+ "secs": 54.66885995864868,
460
+ "gate_secs": 28.582393646240234,
461
+ "samples_per_sec": 219.50338838374816
462
+ },
463
+ "secs": 114.41784977912903,
464
+ "rl_secs_used": 486.9416391849518
465
+ },
466
+ {
467
+ "round": 10,
468
+ "level": 2,
469
+ "k": 24,
470
+ "n_targets": 203,
471
+ "sample_acc": 0.06547619047619048,
472
+ "new_proofs": 48,
473
+ "targets_solved": 949,
474
+ "lstar": 10,
475
+ "ge": {
476
+ "7": 949,
477
+ "8": 449,
478
+ "9": 202,
479
+ "10": 41,
480
+ "11": 1,
481
+ "12": 0,
482
+ "13": 0,
483
+ "14": 0
484
+ },
485
+ "mix_rl_records": 58,
486
+ "sampling": {
487
+ "samples": 4872,
488
+ "secs": 29.055354356765747,
489
+ "gate_secs": 13.798458814620972,
490
+ "samples_per_sec": 167.67993741110647
491
+ },
492
+ "secs": 29.263020753860474,
493
+ "rl_secs_used": 502.40589570999146
494
+ },
495
+ {
496
+ "round": 11,
497
+ "level": 4,
498
+ "k": 24,
499
+ "n_targets": 203,
500
+ "sample_acc": 0.09072249589490969,
501
+ "new_proofs": 51,
502
+ "targets_solved": 983,
503
+ "lstar": 10,
504
+ "ge": {
505
+ "7": 983,
506
+ "8": 473,
507
+ "9": 220,
508
+ "10": 46,
509
+ "11": 1,
510
+ "12": 0,
511
+ "13": 0,
512
+ "14": 0
513
+ },
514
+ "mix_rl_records": 69,
515
+ "sampling": {
516
+ "samples": 4872,
517
+ "secs": 34.9158148765564,
518
+ "gate_secs": 13.667944431304932,
519
+ "samples_per_sec": 139.53562353405698
520
+ },
521
+ "secs": 35.08571791648865,
522
+ "rl_secs_used": 523.8236849308014
523
+ },
524
+ {
525
+ "round": 12,
526
+ "level": 8,
527
+ "k": 24,
528
+ "n_targets": 203,
529
+ "sample_acc": 0.10940065681444992,
530
+ "new_proofs": 65,
531
+ "targets_solved": 1021,
532
+ "lstar": 10,
533
+ "ge": {
534
+ "7": 1021,
535
+ "8": 500,
536
+ "9": 243,
537
+ "10": 50,
538
+ "11": 1,
539
+ "12": 0,
540
+ "13": 0,
541
+ "14": 0
542
+ },
543
+ "mix_rl_records": 77,
544
+ "sampling": {
545
+ "samples": 4872,
546
+ "secs": 53.40554690361023,
547
+ "gate_secs": 12.501071453094482,
548
+ "samples_per_sec": 91.2264789422211
549
+ },
550
+ "secs": 53.56394910812378,
551
+ "rl_secs_used": 564.8865392208099
552
+ },
553
+ {
554
+ "round": 13,
555
+ "level": 1,
556
+ "k": 8,
557
+ "n_targets": 1500,
558
+ "sample_acc": 0.5436666666666666,
559
+ "new_proofs": 261,
560
+ "targets_solved": 1063,
561
+ "lstar": 11,
562
+ "ge": {
563
+ "7": 1063,
564
+ "8": 526,
565
+ "9": 264,
566
+ "10": 59,
567
+ "11": 5,
568
+ "12": 0,
569
+ "13": 0,
570
+ "14": 0
571
+ },
572
+ "mix_rl_records": 1554,
573
+ "sampling": {
574
+ "samples": 12000,
575
+ "secs": 53.831048250198364,
576
+ "gate_secs": 28.652864694595337,
577
+ "samples_per_sec": 222.91967907119067
578
+ },
579
+ "secs": 115.19747161865234,
580
+ "rl_secs_used": 651.4314601421356
581
+ },
582
+ {
583
+ "round": 14,
584
+ "level": 2,
585
+ "k": 24,
586
+ "n_targets": 160,
587
+ "sample_acc": 0.05755208333333333,
588
+ "new_proofs": 37,
589
+ "targets_solved": 1088,
590
+ "lstar": 11,
591
+ "ge": {
592
+ "7": 1088,
593
+ "8": 546,
594
+ "9": 280,
595
+ "10": 64,
596
+ "11": 8,
597
+ "12": 0,
598
+ "13": 0,
599
+ "14": 0
600
+ },
601
+ "mix_rl_records": 52,
602
+ "sampling": {
603
+ "samples": 3840,
604
+ "secs": 23.893316745758057,
605
+ "gate_secs": 10.834507942199707,
606
+ "samples_per_sec": 160.71439728775795
607
+ },
608
+ "secs": 24.05091643333435,
609
+ "rl_secs_used": 664.6475577354431
610
+ },
611
+ {
612
+ "round": 15,
613
+ "level": 4,
614
+ "k": 24,
615
+ "n_targets": 159,
616
+ "sample_acc": 0.0859538784067086,
617
+ "new_proofs": 36,
618
+ "targets_solved": 1108,
619
+ "lstar": 11,
620
+ "ge": {
621
+ "7": 1108,
622
+ "8": 563,
623
+ "9": 293,
624
+ "10": 68,
625
+ "11": 10,
626
+ "12": 0,
627
+ "13": 0,
628
+ "14": 0
629
+ },
630
+ "mix_rl_records": 51,
631
+ "sampling": {
632
+ "samples": 3816,
633
+ "secs": 29.258893728256226,
634
+ "gate_secs": 10.438514947891235,
635
+ "samples_per_sec": 130.42188250319148
636
+ },
637
+ "secs": 29.411959886550903,
638
+ "rl_secs_used": 683.6210010051727
639
+ },
640
+ {
641
+ "round": 16,
642
+ "level": 8,
643
+ "k": 24,
644
+ "n_targets": 159,
645
+ "sample_acc": 0.10770440251572327,
646
+ "new_proofs": 36,
647
+ "targets_solved": 1124,
648
+ "lstar": 11,
649
+ "ge": {
650
+ "7": 1124,
651
+ "8": 577,
652
+ "9": 303,
653
+ "10": 74,
654
+ "11": 12,
655
+ "12": 1,
656
+ "13": 0,
657
+ "14": 0
658
+ },
659
+ "mix_rl_records": 62,
660
+ "sampling": {
661
+ "samples": 3816,
662
+ "secs": 37.074631452560425,
663
+ "gate_secs": 10.293965339660645,
664
+ "samples_per_sec": 102.92752349764658
665
+ },
666
+ "secs": 37.23300504684448,
667
+ "rl_secs_used": 710.5600562095642
668
+ },
669
+ {
670
+ "round": 17,
671
+ "level": 1,
672
+ "k": 8,
673
+ "n_targets": 1500,
674
+ "sample_acc": 0.6005,
675
+ "new_proofs": 335,
676
+ "targets_solved": 1138,
677
+ "lstar": 11,
678
+ "ge": {
679
+ "7": 1138,
680
+ "8": 587,
681
+ "9": 311,
682
+ "10": 77,
683
+ "11": 12,
684
+ "12": 1,
685
+ "13": 0,
686
+ "14": 0
687
+ },
688
+ "mix_rl_records": 1884,
689
+ "sampling": {
690
+ "samples": 12000,
691
+ "secs": 50.913803815841675,
692
+ "gate_secs": 28.240092277526855,
693
+ "samples_per_sec": 235.69246649503404
694
+ },
695
+ "secs": 113.66210722923279,
696
+ "rl_secs_used": 795.9823000431061
697
+ },
698
+ {
699
+ "round": 18,
700
+ "level": 2,
701
+ "k": 24,
702
+ "n_targets": 133,
703
+ "sample_acc": 0.05231829573934837,
704
+ "new_proofs": 17,
705
+ "targets_solved": 1148,
706
+ "lstar": 11,
707
+ "ge": {
708
+ "7": 1148,
709
+ "8": 594,
710
+ "9": 316,
711
+ "10": 80,
712
+ "11": 13,
713
+ "12": 2,
714
+ "13": 0,
715
+ "14": 0
716
+ },
717
+ "mix_rl_records": 36,
718
+ "sampling": {
719
+ "samples": 3192,
720
+ "secs": 17.69075083732605,
721
+ "gate_secs": 7.808672189712524,
722
+ "samples_per_sec": 180.43326873753367
723
+ },
724
+ "secs": 17.811609506607056,
725
+ "rl_secs_used": 805.984983921051
726
+ },
727
+ {
728
+ "round": 19,
729
+ "level": 4,
730
+ "k": 24,
731
+ "n_targets": 133,
732
+ "sample_acc": 0.07769423558897243,
733
+ "new_proofs": 27,
734
+ "targets_solved": 1160,
735
+ "lstar": 11,
736
+ "ge": {
737
+ "7": 1160,
738
+ "8": 605,
739
+ "9": 326,
740
+ "10": 87,
741
+ "11": 17,
742
+ "12": 2,
743
+ "13": 0,
744
+ "14": 0
745
+ },
746
+ "mix_rl_records": 42,
747
+ "sampling": {
748
+ "samples": 3192,
749
+ "secs": 24.773690462112427,
750
+ "gate_secs": 7.864348411560059,
751
+ "samples_per_sec": 128.8463664661378
752
+ },
753
+ "secs": 24.896905660629272,
754
+ "rl_secs_used": 823.0175387859344
755
+ },
756
+ {
757
+ "round": 20,
758
+ "level": 8,
759
+ "k": 24,
760
+ "n_targets": 132,
761
+ "sample_acc": 0.08017676767676768,
762
+ "new_proofs": 32,
763
+ "targets_solved": 1173,
764
+ "lstar": 11,
765
+ "ge": {
766
+ "7": 1173,
767
+ "8": 616,
768
+ "9": 335,
769
+ "10": 90,
770
+ "11": 18,
771
+ "12": 2,
772
+ "13": 0,
773
+ "14": 0
774
+ },
775
+ "mix_rl_records": 44,
776
+ "sampling": {
777
+ "samples": 3168,
778
+ "secs": 31.244083404541016,
779
+ "gate_secs": 7.392596483230591,
780
+ "samples_per_sec": 101.39519725963741
781
+ },
782
+ "secs": 31.361677169799805,
783
+ "rl_secs_used": 846.9866254329681
784
+ },
785
+ {
786
+ "round": 21,
787
+ "level": 1,
788
+ "k": 8,
789
+ "n_targets": 1500,
790
+ "sample_acc": 0.6416666666666667,
791
+ "new_proofs": 399,
792
+ "targets_solved": 1176,
793
+ "lstar": 11,
794
+ "ge": {
795
+ "7": 1176,
796
+ "8": 618,
797
+ "9": 337,
798
+ "10": 91,
799
+ "11": 19,
800
+ "12": 2,
801
+ "13": 0,
802
+ "14": 0
803
+ },
804
+ "mix_rl_records": 2199,
805
+ "sampling": {
806
+ "samples": 12000,
807
+ "secs": 49.32094049453735,
808
+ "gate_secs": 28.645761251449585,
809
+ "samples_per_sec": 243.30436280567451
810
+ },
811
+ "secs": 111.92527604103088,
812
+ "rl_secs_used": 930.266375541687
813
+ },
814
+ {
815
+ "round": 22,
816
+ "level": 2,
817
+ "k": 24,
818
+ "n_targets": 117,
819
+ "sample_acc": 0.04202279202279202,
820
+ "new_proofs": 14,
821
+ "targets_solved": 1179,
822
+ "lstar": 11,
823
+ "ge": {
824
+ "7": 1179,
825
+ "8": 621,
826
+ "9": 339,
827
+ "10": 91,
828
+ "11": 19,
829
+ "12": 2,
830
+ "13": 0,
831
+ "14": 0
832
+ },
833
+ "mix_rl_records": 26,
834
+ "sampling": {
835
+ "samples": 2808,
836
+ "secs": 17.28331446647644,
837
+ "gate_secs": 6.623600482940674,
838
+ "samples_per_sec": 162.4688369494482
839
+ },
840
+ "secs": 17.384339809417725,
841
+ "rl_secs_used": 941.0268788337708
842
+ },
843
+ {
844
+ "round": 23,
845
+ "level": 4,
846
+ "k": 24,
847
+ "n_targets": 117,
848
+ "sample_acc": 0.06588319088319089,
849
+ "new_proofs": 14,
850
+ "targets_solved": 1187,
851
+ "lstar": 11,
852
+ "ge": {
853
+ "7": 1187,
854
+ "8": 625,
855
+ "9": 343,
856
+ "10": 93,
857
+ "11": 20,
858
+ "12": 2,
859
+ "13": 0,
860
+ "14": 0
861
+ },
862
+ "mix_rl_records": 25,
863
+ "sampling": {
864
+ "samples": 2808,
865
+ "secs": 21.400601863861084,
866
+ "gate_secs": 6.9875688552856445,
867
+ "samples_per_sec": 131.21126302255232
868
+ },
869
+ "secs": 21.50975012779236,
870
+ "rl_secs_used": 955.549069404602
871
+ },
872
+ {
873
+ "round": 24,
874
+ "level": 8,
875
+ "k": 24,
876
+ "n_targets": 116,
877
+ "sample_acc": 0.04274425287356322,
878
+ "new_proofs": 8,
879
+ "targets_solved": 1193,
880
+ "lstar": 11,
881
+ "ge": {
882
+ "7": 1193,
883
+ "8": 630,
884
+ "9": 348,
885
+ "10": 94,
886
+ "11": 20,
887
+ "12": 2,
888
+ "13": 0,
889
+ "14": 0
890
+ },
891
+ "mix_rl_records": 14,
892
+ "sampling": {
893
+ "samples": 2784,
894
+ "secs": 30.261083841323853,
895
+ "gate_secs": 6.381618976593018,
896
+ "samples_per_sec": 91.99934855599034
897
+ },
898
+ "secs": 30.357787370681763,
899
+ "rl_secs_used": 979.5252232551575
900
+ },
901
+ {
902
+ "round": 25,
903
+ "level": 1,
904
+ "k": 8,
905
+ "n_targets": 1500,
906
+ "sample_acc": 0.6465,
907
+ "new_proofs": 292,
908
+ "targets_solved": 1197,
909
+ "lstar": 11,
910
+ "ge": {
911
+ "7": 1197,
912
+ "8": 632,
913
+ "9": 350,
914
+ "10": 95,
915
+ "11": 20,
916
+ "12": 2,
917
+ "13": 0,
918
+ "14": 0
919
+ },
920
+ "mix_rl_records": 2275,
921
+ "sampling": {
922
+ "samples": 12000,
923
+ "secs": 50.54125380516052,
924
+ "gate_secs": 29.25943636894226,
925
+ "samples_per_sec": 237.42980429928983
926
+ },
927
+ "secs": 115.70601153373718,
928
+ "rl_secs_used": 1065.972058057785
929
+ },
930
+ {
931
+ "round": 26,
932
+ "level": 2,
933
+ "k": 24,
934
+ "n_targets": 114,
935
+ "sample_acc": 0.05116959064327485,
936
+ "new_proofs": 14,
937
+ "targets_solved": 1202,
938
+ "lstar": 11,
939
+ "ge": {
940
+ "7": 1202,
941
+ "8": 636,
942
+ "9": 353,
943
+ "10": 96,
944
+ "11": 20,
945
+ "12": 2,
946
+ "13": 0,
947
+ "14": 0
948
+ },
949
+ "mix_rl_records": 37,
950
+ "sampling": {
951
+ "samples": 2736,
952
+ "secs": 15.594995260238647,
953
+ "gate_secs": 6.374717950820923,
954
+ "samples_per_sec": 175.44089974658522
955
+ },
956
+ "secs": 15.696367979049683,
957
+ "rl_secs_used": 1075.2934577465057
958
+ },
959
+ {
960
+ "round": 27,
961
+ "level": 4,
962
+ "k": 24,
963
+ "n_targets": 113,
964
+ "sample_acc": 0.09070796460176991,
965
+ "new_proofs": 9,
966
+ "targets_solved": 1207,
967
+ "lstar": 11,
968
+ "ge": {
969
+ "7": 1207,
970
+ "8": 639,
971
+ "9": 355,
972
+ "10": 98,
973
+ "11": 20,
974
+ "12": 2,
975
+ "13": 0,
976
+ "14": 0
977
+ },
978
+ "mix_rl_records": 24,
979
+ "sampling": {
980
+ "samples": 2712,
981
+ "secs": 22.035794258117676,
982
+ "gate_secs": 5.325361967086792,
983
+ "samples_per_sec": 123.0724868925901
984
+ },
985
+ "secs": 22.127794981002808,
986
+ "rl_secs_used": 1092.0958905220032
987
+ },
988
+ {
989
+ "round": 28,
990
+ "level": 8,
991
+ "k": 24,
992
+ "n_targets": 113,
993
+ "sample_acc": 0.06858407079646017,
994
+ "new_proofs": 20,
995
+ "targets_solved": 1215,
996
+ "lstar": 11,
997
+ "ge": {
998
+ "7": 1215,
999
+ "8": 643,
1000
+ "9": 357,
1001
+ "10": 98,
1002
+ "11": 20,
1003
+ "12": 2,
1004
+ "13": 0,
1005
+ "14": 0
1006
+ },
1007
+ "mix_rl_records": 37,
1008
+ "sampling": {
1009
+ "samples": 2712,
1010
+ "secs": 36.153486490249634,
1011
+ "gate_secs": 5.900385141372681,
1012
+ "samples_per_sec": 75.01351220252047
1013
+ },
1014
+ "secs": 36.24789643287659,
1015
+ "rl_secs_used": 1122.4434049129486
1016
+ },
1017
+ {
1018
+ "round": 29,
1019
+ "level": 1,
1020
+ "k": 8,
1021
+ "n_targets": 1500,
1022
+ "sample_acc": 0.6625,
1023
+ "new_proofs": 325,
1024
+ "targets_solved": 1216,
1025
+ "lstar": 11,
1026
+ "ge": {
1027
+ "7": 1216,
1028
+ "8": 643,
1029
+ "9": 357,
1030
+ "10": 98,
1031
+ "11": 20,
1032
+ "12": 2,
1033
+ "13": 0,
1034
+ "14": 0
1035
+ },
1036
+ "mix_rl_records": 2513,
1037
+ "sampling": {
1038
+ "samples": 12000,
1039
+ "secs": 52.666664361953735,
1040
+ "gate_secs": 27.859134912490845,
1041
+ "samples_per_sec": 227.84811123654092
1042
+ },
1043
+ "secs": 121.3718490600586,
1044
+ "rl_secs_used": 1215.9563705921173
1045
+ },
1046
+ {
1047
+ "round": 30,
1048
+ "level": 2,
1049
+ "k": 24,
1050
+ "n_targets": 104,
1051
+ "sample_acc": 0.02403846153846154,
1052
+ "new_proofs": 7,
1053
+ "targets_solved": 1219,
1054
+ "lstar": 11,
1055
+ "ge": {
1056
+ "7": 1219,
1057
+ "8": 645,
1058
+ "9": 359,
1059
+ "10": 99,
1060
+ "11": 21,
1061
+ "12": 3,
1062
+ "13": 0,
1063
+ "14": 0
1064
+ },
1065
+ "mix_rl_records": 12,
1066
+ "sampling": {
1067
+ "samples": 2496,
1068
+ "secs": 13.99475622177124,
1069
+ "gate_secs": 5.13289737701416,
1070
+ "samples_per_sec": 178.35251721763075
1071
+ },
1072
+ "secs": 14.06341290473938,
1073
+ "rl_secs_used": 1224.8866262435913
1074
+ },
1075
+ {
1076
+ "round": 31,
1077
+ "level": 4,
1078
+ "k": 24,
1079
+ "n_targets": 104,
1080
+ "sample_acc": 0.041666666666666664,
1081
+ "new_proofs": 3,
1082
+ "targets_solved": 1220,
1083
+ "lstar": 11,
1084
+ "ge": {
1085
+ "7": 1220,
1086
+ "8": 645,
1087
+ "9": 359,
1088
+ "10": 99,
1089
+ "11": 21,
1090
+ "12": 3,
1091
+ "13": 0,
1092
+ "14": 0
1093
+ },
1094
+ "mix_rl_records": 14,
1095
+ "sampling": {
1096
+ "samples": 2496,
1097
+ "secs": 19.101277828216553,
1098
+ "gate_secs": 5.16304349899292,
1099
+ "samples_per_sec": 130.67188606161676
1100
+ },
1101
+ "secs": 19.188413381576538,
1102
+ "rl_secs_used": 1238.9119930267334
1103
+ },
1104
+ {
1105
+ "round": 32,
1106
+ "level": 8,
1107
+ "k": 24,
1108
+ "n_targets": 104,
1109
+ "sample_acc": 0.12139423076923077,
1110
+ "new_proofs": 21,
1111
+ "targets_solved": 1226,
1112
+ "lstar": 11,
1113
+ "ge": {
1114
+ "7": 1226,
1115
+ "8": 648,
1116
+ "9": 361,
1117
+ "10": 100,
1118
+ "11": 21,
1119
+ "12": 3,
1120
+ "13": 0,
1121
+ "14": 0
1122
+ },
1123
+ "mix_rl_records": 46,
1124
+ "sampling": {
1125
+ "samples": 2496,
1126
+ "secs": 31.205972909927368,
1127
+ "gate_secs": 5.180824279785156,
1128
+ "samples_per_sec": 79.98468777770304
1129
+ },
1130
+ "secs": 31.295889616012573,
1131
+ "rl_secs_used": 1265.0270555019379
1132
+ },
1133
+ {
1134
+ "round": 33,
1135
+ "level": 1,
1136
+ "k": 8,
1137
+ "n_targets": 1500,
1138
+ "sample_acc": 0.6650833333333334,
1139
+ "new_proofs": 313,
1140
+ "targets_solved": 1228,
1141
+ "lstar": 11,
1142
+ "ge": {
1143
+ "7": 1228,
1144
+ "8": 650,
1145
+ "9": 363,
1146
+ "10": 101,
1147
+ "11": 21,
1148
+ "12": 3,
1149
+ "13": 0,
1150
+ "14": 0
1151
+ },
1152
+ "mix_rl_records": 2544,
1153
+ "sampling": {
1154
+ "samples": 12000,
1155
+ "secs": 53.99081754684448,
1156
+ "gate_secs": 30.00288200378418,
1157
+ "samples_per_sec": 222.26001652203811
1158
+ },
1159
+ "secs": 122.35320520401001,
1160
+ "rl_secs_used": 1357.3776247501373
1161
+ },
1162
+ {
1163
+ "round": 34,
1164
+ "level": 2,
1165
+ "k": 24,
1166
+ "n_targets": 101,
1167
+ "sample_acc": 0.06105610561056106,
1168
+ "new_proofs": 13,
1169
+ "targets_solved": 1234,
1170
+ "lstar": 11,
1171
+ "ge": {
1172
+ "7": 1234,
1173
+ "8": 655,
1174
+ "9": 363,
1175
+ "10": 101,
1176
+ "11": 21,
1177
+ "12": 3,
1178
+ "13": 0,
1179
+ "14": 0
1180
+ },
1181
+ "mix_rl_records": 32,
1182
+ "sampling": {
1183
+ "samples": 2424,
1184
+ "secs": 14.191061019897461,
1185
+ "gate_secs": 5.057612657546997,
1186
+ "samples_per_sec": 170.81175231374735
1187
+ },
1188
+ "secs": 14.27733302116394,
1189
+ "rl_secs_used": 1366.597109079361
1190
+ },
1191
+ {
1192
+ "round": 35,
1193
+ "level": 4,
1194
+ "k": 24,
1195
+ "n_targets": 101,
1196
+ "sample_acc": 0.06435643564356436,
1197
+ "new_proofs": 8,
1198
+ "targets_solved": 1237,
1199
+ "lstar": 11,
1200
+ "ge": {
1201
+ "7": 1237,
1202
+ "8": 656,
1203
+ "9": 364,
1204
+ "10": 101,
1205
+ "11": 21,
1206
+ "12": 3,
1207
+ "13": 0,
1208
+ "14": 0
1209
+ },
1210
+ "mix_rl_records": 17,
1211
+ "sampling": {
1212
+ "samples": 2424,
1213
+ "secs": 18.872806549072266,
1214
+ "gate_secs": 5.000371932983398,
1215
+ "samples_per_sec": 128.43876684144558
1216
+ },
1217
+ "secs": 18.95445942878723,
1218
+ "rl_secs_used": 1380.5512008666992
1219
+ },
1220
+ {
1221
+ "round": 36,
1222
+ "level": 8,
1223
+ "k": 24,
1224
+ "n_targets": 101,
1225
+ "sample_acc": 0.09570957095709572,
1226
+ "new_proofs": 12,
1227
+ "targets_solved": 1238,
1228
+ "lstar": 11,
1229
+ "ge": {
1230
+ "7": 1238,
1231
+ "8": 657,
1232
+ "9": 365,
1233
+ "10": 101,
1234
+ "11": 21,
1235
+ "12": 3,
1236
+ "13": 0,
1237
+ "14": 0
1238
+ },
1239
+ "mix_rl_records": 33,
1240
+ "sampling": {
1241
+ "samples": 2424,
1242
+ "secs": 27.07240104675293,
1243
+ "gate_secs": 5.202834129333496,
1244
+ "samples_per_sec": 89.53768067390295
1245
+ },
1246
+ "secs": 27.161224842071533,
1247
+ "rl_secs_used": 1402.5095844268799
1248
+ },
1249
+ {
1250
+ "round": 37,
1251
+ "level": 1,
1252
+ "k": 8,
1253
+ "n_targets": 1500,
1254
+ "sample_acc": 0.6971666666666667,
1255
+ "new_proofs": 401,
1256
+ "targets_solved": 1242,
1257
+ "lstar": 11,
1258
+ "ge": {
1259
+ "7": 1242,
1260
+ "8": 661,
1261
+ "9": 369,
1262
+ "10": 103,
1263
+ "11": 22,
1264
+ "12": 3,
1265
+ "13": 0,
1266
+ "14": 0
1267
+ },
1268
+ "mix_rl_records": 2868,
1269
+ "sampling": {
1270
+ "samples": 12000,
1271
+ "secs": 55.002439737319946,
1272
+ "gate_secs": 29.064289331436157,
1273
+ "samples_per_sec": 218.17214031431095
1274
+ },
1275
+ "secs": 122.61763095855713,
1276
+ "rl_secs_used": 1496.063163280487
1277
+ },
1278
+ {
1279
+ "round": 38,
1280
+ "level": 2,
1281
+ "k": 24,
1282
+ "n_targets": 92,
1283
+ "sample_acc": 0.049365942028985504,
1284
+ "new_proofs": 3,
1285
+ "targets_solved": 1245,
1286
+ "lstar": 11,
1287
+ "ge": {
1288
+ "7": 1245,
1289
+ "8": 664,
1290
+ "9": 369,
1291
+ "10": 103,
1292
+ "11": 22,
1293
+ "12": 3,
1294
+ "13": 0,
1295
+ "14": 0
1296
+ },
1297
+ "mix_rl_records": 16,
1298
+ "sampling": {
1299
+ "samples": 2208,
1300
+ "secs": 13.646290302276611,
1301
+ "gate_secs": 4.990728139877319,
1302
+ "samples_per_sec": 161.802215187496
1303
+ },
1304
+ "secs": 13.724525690078735,
1305
+ "rl_secs_used": 1504.7967267036438
1306
+ },
1307
+ {
1308
+ "round": 39,
1309
+ "level": 4,
1310
+ "k": 24,
1311
+ "n_targets": 92,
1312
+ "sample_acc": 0.021286231884057972,
1313
+ "new_proofs": 5,
1314
+ "targets_solved": 1249,
1315
+ "lstar": 11,
1316
+ "ge": {
1317
+ "7": 1249,
1318
+ "8": 668,
1319
+ "9": 371,
1320
+ "10": 103,
1321
+ "11": 22,
1322
+ "12": 3,
1323
+ "13": 0,
1324
+ "14": 0
1325
+ },
1326
+ "mix_rl_records": 8,
1327
+ "sampling": {
1328
+ "samples": 2208,
1329
+ "secs": 18.13883686065674,
1330
+ "gate_secs": 4.583757400512695,
1331
+ "samples_per_sec": 121.72776109967488
1332
+ },
1333
+ "secs": 18.20617389678955,
1334
+ "rl_secs_used": 1518.419132232666
1335
+ },
1336
+ {
1337
+ "round": 40,
1338
+ "level": 8,
1339
+ "k": 24,
1340
+ "n_targets": 91,
1341
+ "sample_acc": 0.06913919413919414,
1342
+ "new_proofs": 18,
1343
+ "targets_solved": 1257,
1344
+ "lstar": 11,
1345
+ "ge": {
1346
+ "7": 1257,
1347
+ "8": 674,
1348
+ "9": 375,
1349
+ "10": 105,
1350
+ "11": 22,
1351
+ "12": 3,
1352
+ "13": 0,
1353
+ "14": 0
1354
+ },
1355
+ "mix_rl_records": 32,
1356
+ "sampling": {
1357
+ "samples": 2184,
1358
+ "secs": 26.643465280532837,
1359
+ "gate_secs": 4.7565999031066895,
1360
+ "samples_per_sec": 81.97131930866925
1361
+ },
1362
+ "secs": 26.72255539894104,
1363
+ "rl_secs_used": 1540.3850934505463
1364
+ },
1365
+ {
1366
+ "round": 41,
1367
+ "level": 1,
1368
+ "k": 8,
1369
+ "n_targets": 1500,
1370
+ "sample_acc": 0.703,
1371
+ "new_proofs": 335,
1372
+ "targets_solved": 1259,
1373
+ "lstar": 11,
1374
+ "ge": {
1375
+ "7": 1259,
1376
+ "8": 676,
1377
+ "9": 375,
1378
+ "10": 105,
1379
+ "11": 22,
1380
+ "12": 3,
1381
+ "13": 0,
1382
+ "14": 0
1383
+ },
1384
+ "mix_rl_records": 2936,
1385
+ "sampling": {
1386
+ "samples": 12000,
1387
+ "secs": 53.42039132118225,
1388
+ "gate_secs": 29.173763275146484,
1389
+ "samples_per_sec": 224.6333226548597
1390
+ },
1391
+ "secs": 122.11814832687378,
1392
+ "rl_secs_used": 1633.3296978473663
1393
+ },
1394
+ {
1395
+ "round": 42,
1396
+ "level": 2,
1397
+ "k": 24,
1398
+ "n_targets": 88,
1399
+ "sample_acc": 0.06202651515151515,
1400
+ "new_proofs": 4,
1401
+ "targets_solved": 1259,
1402
+ "lstar": 11,
1403
+ "ge": {
1404
+ "7": 1259,
1405
+ "8": 676,
1406
+ "9": 375,
1407
+ "10": 105,
1408
+ "11": 22,
1409
+ "12": 3,
1410
+ "13": 0,
1411
+ "14": 0
1412
+ },
1413
+ "mix_rl_records": 22,
1414
+ "sampling": {
1415
+ "samples": 2112,
1416
+ "secs": 12.783551692962646,
1417
+ "gate_secs": 4.385697364807129,
1418
+ "samples_per_sec": 165.21230177077138
1419
+ },
1420
+ "secs": 12.857568740844727,
1421
+ "rl_secs_used": 1641.8013417720795
1422
+ },
1423
+ {
1424
+ "round": 43,
1425
+ "level": 4,
1426
+ "k": 24,
1427
+ "n_targets": 88,
1428
+ "sample_acc": 0.048295454545454544,
1429
+ "new_proofs": 2,
1430
+ "targets_solved": 1260,
1431
+ "lstar": 11,
1432
+ "ge": {
1433
+ "7": 1260,
1434
+ "8": 677,
1435
+ "9": 375,
1436
+ "10": 105,
1437
+ "11": 22,
1438
+ "12": 3,
1439
+ "13": 0,
1440
+ "14": 0
1441
+ },
1442
+ "mix_rl_records": 10,
1443
+ "sampling": {
1444
+ "samples": 2112,
1445
+ "secs": 18.436086177825928,
1446
+ "gate_secs": 4.061391592025757,
1447
+ "samples_per_sec": 114.5579370604275
1448
+ },
1449
+ "secs": 18.501312732696533,
1450
+ "rl_secs_used": 1656.241271018982
1451
+ },
1452
+ {
1453
+ "round": 44,
1454
+ "level": 8,
1455
+ "k": 24,
1456
+ "n_targets": 87,
1457
+ "sample_acc": 0.07375478927203065,
1458
+ "new_proofs": 13,
1459
+ "targets_solved": 1264,
1460
+ "lstar": 11,
1461
+ "ge": {
1462
+ "7": 1264,
1463
+ "8": 679,
1464
+ "9": 377,
1465
+ "10": 107,
1466
+ "11": 24,
1467
+ "12": 4,
1468
+ "13": 0,
1469
+ "14": 0
1470
+ },
1471
+ "mix_rl_records": 28,
1472
+ "sampling": {
1473
+ "samples": 2088,
1474
+ "secs": 31.791592359542847,
1475
+ "gate_secs": 4.119357585906982,
1476
+ "samples_per_sec": 65.67774197611864
1477
+ },
1478
+ "secs": 31.8604097366333,
1479
+ "rl_secs_used": 1683.9823250770569
1480
+ }
1481
+ ],
1482
+ "secs": 2328.132572889328,
1483
+ "rl_secs_used": 1755.4754884243011,
1484
+ "gate_secs": 572.6251745223999,
1485
+ "targets": 1500,
1486
+ "stats": {
1487
+ "rounds": 11,
1488
+ "first_loops": 1,
1489
+ "k_first": 8,
1490
+ "retry_loops": [
1491
+ 2,
1492
+ 4,
1493
+ 8
1494
+ ],
1495
+ "k_retry": 24,
1496
+ "retries": [
1497
+ {
1498
+ "round": 1,
1499
+ "loops": 2,
1500
+ "targets": 489,
1501
+ "cracked": 33
1502
+ },
1503
+ {
1504
+ "round": 1,
1505
+ "loops": 4,
1506
+ "targets": 489,
1507
+ "cracked": 53
1508
+ },
1509
+ {
1510
+ "round": 1,
1511
+ "loops": 8,
1512
+ "targets": 488,
1513
+ "cracked": 51
1514
+ },
1515
+ {
1516
+ "round": 2,
1517
+ "loops": 2,
1518
+ "targets": 278,
1519
+ "cracked": 50
1520
+ },
1521
+ {
1522
+ "round": 2,
1523
+ "loops": 4,
1524
+ "targets": 278,
1525
+ "cracked": 66
1526
+ },
1527
+ {
1528
+ "round": 2,
1529
+ "loops": 8,
1530
+ "targets": 277,
1531
+ "cracked": 71
1532
+ },
1533
+ {
1534
+ "round": 3,
1535
+ "loops": 2,
1536
+ "targets": 203,
1537
+ "cracked": 36
1538
+ },
1539
+ {
1540
+ "round": 3,
1541
+ "loops": 4,
1542
+ "targets": 203,
1543
+ "cracked": 47
1544
+ },
1545
+ {
1546
+ "round": 3,
1547
+ "loops": 8,
1548
+ "targets": 203,
1549
+ "cracked": 47
1550
+ },
1551
+ {
1552
+ "round": 4,
1553
+ "loops": 2,
1554
+ "targets": 160,
1555
+ "cracked": 37
1556
+ },
1557
+ {
1558
+ "round": 4,
1559
+ "loops": 4,
1560
+ "targets": 159,
1561
+ "cracked": 32
1562
+ },
1563
+ {
1564
+ "round": 4,
1565
+ "loops": 8,
1566
+ "targets": 159,
1567
+ "cracked": 33
1568
+ },
1569
+ {
1570
+ "round": 5,
1571
+ "loops": 2,
1572
+ "targets": 133,
1573
+ "cracked": 22
1574
+ },
1575
+ {
1576
+ "round": 5,
1577
+ "loops": 4,
1578
+ "targets": 133,
1579
+ "cracked": 24
1580
+ },
1581
+ {
1582
+ "round": 5,
1583
+ "loops": 8,
1584
+ "targets": 132,
1585
+ "cracked": 23
1586
+ },
1587
+ {
1588
+ "round": 6,
1589
+ "loops": 2,
1590
+ "targets": 117,
1591
+ "cracked": 13
1592
+ },
1593
+ {
1594
+ "round": 6,
1595
+ "loops": 4,
1596
+ "targets": 117,
1597
+ "cracked": 19
1598
+ },
1599
+ {
1600
+ "round": 6,
1601
+ "loops": 8,
1602
+ "targets": 116,
1603
+ "cracked": 11
1604
+ },
1605
+ {
1606
+ "round": 7,
1607
+ "loops": 2,
1608
+ "targets": 114,
1609
+ "cracked": 17
1610
+ },
1611
+ {
1612
+ "round": 7,
1613
+ "loops": 4,
1614
+ "targets": 113,
1615
+ "cracked": 18
1616
+ },
1617
+ {
1618
+ "round": 7,
1619
+ "loops": 8,
1620
+ "targets": 113,
1621
+ "cracked": 19
1622
+ },
1623
+ {
1624
+ "round": 8,
1625
+ "loops": 2,
1626
+ "targets": 104,
1627
+ "cracked": 8
1628
+ },
1629
+ {
1630
+ "round": 8,
1631
+ "loops": 4,
1632
+ "targets": 104,
1633
+ "cracked": 8
1634
+ },
1635
+ {
1636
+ "round": 8,
1637
+ "loops": 8,
1638
+ "targets": 104,
1639
+ "cracked": 20
1640
+ },
1641
+ {
1642
+ "round": 9,
1643
+ "loops": 2,
1644
+ "targets": 101,
1645
+ "cracked": 17
1646
+ },
1647
+ {
1648
+ "round": 9,
1649
+ "loops": 4,
1650
+ "targets": 101,
1651
+ "cracked": 11
1652
+ },
1653
+ {
1654
+ "round": 9,
1655
+ "loops": 8,
1656
+ "targets": 101,
1657
+ "cracked": 13
1658
+ },
1659
+ {
1660
+ "round": 10,
1661
+ "loops": 2,
1662
+ "targets": 92,
1663
+ "cracked": 11
1664
+ },
1665
+ {
1666
+ "round": 10,
1667
+ "loops": 4,
1668
+ "targets": 92,
1669
+ "cracked": 6
1670
+ },
1671
+ {
1672
+ "round": 10,
1673
+ "loops": 8,
1674
+ "targets": 91,
1675
+ "cracked": 15
1676
+ },
1677
+ {
1678
+ "round": 11,
1679
+ "loops": 2,
1680
+ "targets": 88,
1681
+ "cracked": 8
1682
+ },
1683
+ {
1684
+ "round": 11,
1685
+ "loops": 4,
1686
+ "targets": 88,
1687
+ "cracked": 6
1688
+ },
1689
+ {
1690
+ "round": 11,
1691
+ "loops": 8,
1692
+ "targets": 87,
1693
+ "cracked": 13
1694
+ }
1695
+ ],
1696
+ "ft_records": 76800,
1697
+ "ft_steps": [
1698
+ 582,
1699
+ 642,
1700
+ 665,
1701
+ 681,
1702
+ 701,
1703
+ 710,
1704
+ 730,
1705
+ 728,
1706
+ 738,
1707
+ 744,
1708
+ 752
1709
+ ]
1710
+ }
1711
+ },
1712
+ "eval": {
1713
+ "level": 8,
1714
+ "attempts_per_theorem": 64,
1715
+ "solved": 765,
1716
+ "n": 1108,
1717
+ "solved_ge7": 765,
1718
+ "lstar": 12,
1719
+ "ge": {
1720
+ "7": 765,
1721
+ "8": 647,
1722
+ "9": 516,
1723
+ "10": 131,
1724
+ "11": 33,
1725
+ "12": 8,
1726
+ "13": 2,
1727
+ "14": 0
1728
+ },
1729
+ "by_bin": {
1730
+ "7": {
1731
+ "solved": 118,
1732
+ "n": 150,
1733
+ "rate": 0.7866666666666666,
1734
+ "ci": [
1735
+ 0.7143779835916132,
1736
+ 0.8446385632981747
1737
+ ]
1738
+ },
1739
+ "8": {
1740
+ "solved": 131,
1741
+ "n": 152,
1742
+ "rate": 0.8618421052631579,
1743
+ "ci": [
1744
+ 0.7980159638451529,
1745
+ 0.907828945344691
1746
+ ]
1747
+ },
1748
+ "9": {
1749
+ "solved": 385,
1750
+ "n": 477,
1751
+ "rate": 0.8071278825995807,
1752
+ "ci": [
1753
+ 0.7693224926693184,
1754
+ 0.8400257833533883
1755
+ ]
1756
+ },
1757
+ "10": {
1758
+ "solved": 98,
1759
+ "n": 219,
1760
+ "rate": 0.4474885844748858,
1761
+ "ci": [
1762
+ 0.3831016383475324,
1763
+ 0.5136860350491762
1764
+ ]
1765
+ },
1766
+ "11": {
1767
+ "solved": 25,
1768
+ "n": 49,
1769
+ "rate": 0.5102040816326531,
1770
+ "ci": [
1771
+ 0.37467331533409215,
1772
+ 0.6442511684059913
1773
+ ]
1774
+ },
1775
+ "12": {
1776
+ "solved": 6,
1777
+ "n": 54,
1778
+ "rate": 0.1111111111111111,
1779
+ "ci": [
1780
+ 0.05192951817556753,
1781
+ 0.2219494893207673
1782
+ ]
1783
+ },
1784
+ "13": {
1785
+ "solved": 2,
1786
+ "n": 5,
1787
+ "rate": 0.4,
1788
+ "ci": [
1789
+ 0.1176182311592533,
1790
+ 0.769280067791163
1791
+ ]
1792
+ },
1793
+ "14": {
1794
+ "solved": 0,
1795
+ "n": 2,
1796
+ "rate": 0.0,
1797
+ "ci": [
1798
+ 0.0,
1799
+ 0.6576280471103807
1800
+ ]
1801
+ }
1802
+ },
1803
+ "sample_acc": 0.48324684115523464,
1804
+ "max_pruned": 15,
1805
+ "sampling": {
1806
+ "samples": 70912,
1807
+ "secs": 770.5194206237793,
1808
+ "gate_secs": 110.5384566783905,
1809
+ "samples_per_sec": 92.03142464935239
1810
+ },
1811
+ "heldout_greedy": 0.9875,
1812
+ "diag": {
1813
+ "level": 2,
1814
+ "attempts_per_theorem": 64,
1815
+ "solved": 743,
1816
+ "n": 1108,
1817
+ "solved_ge7": 743,
1818
+ "lstar": 12,
1819
+ "ge": {
1820
+ "7": 743,
1821
+ "8": 630,
1822
+ "9": 505,
1823
+ "10": 124,
1824
+ "11": 32,
1825
+ "12": 9,
1826
+ "13": 2,
1827
+ "14": 0
1828
+ },
1829
+ "by_bin": {
1830
+ "7": {
1831
+ "solved": 113,
1832
+ "n": 150,
1833
+ "rate": 0.7533333333333333,
1834
+ "ci": [
1835
+ 0.6785952428848271,
1836
+ 0.8154193799480086
1837
+ ]
1838
+ },
1839
+ "8": {
1840
+ "solved": 125,
1841
+ "n": 152,
1842
+ "rate": 0.8223684210526315,
1843
+ "ci": [
1844
+ 0.7538901459847236,
1845
+ 0.87495350038441
1846
+ ]
1847
+ },
1848
+ "9": {
1849
+ "solved": 381,
1850
+ "n": 477,
1851
+ "rate": 0.7987421383647799,
1852
+ "ci": [
1853
+ 0.7604387413371223,
1854
+ 0.8322720390945209
1855
+ ]
1856
+ },
1857
+ "10": {
1858
+ "solved": 92,
1859
+ "n": 219,
1860
+ "rate": 0.4200913242009132,
1861
+ "ci": [
1862
+ 0.3566490808626235,
1863
+ 0.4862886830019332
1864
+ ]
1865
+ },
1866
+ "11": {
1867
+ "solved": 23,
1868
+ "n": 49,
1869
+ "rate": 0.46938775510204084,
1870
+ "ci": [
1871
+ 0.33703281897894777,
1872
+ 0.6061937298008014
1873
+ ]
1874
+ },
1875
+ "12": {
1876
+ "solved": 7,
1877
+ "n": 54,
1878
+ "rate": 0.12962962962962962,
1879
+ "ci": [
1880
+ 0.06423639229487982,
1881
+ 0.2442198053206771
1882
+ ]
1883
+ },
1884
+ "13": {
1885
+ "solved": 2,
1886
+ "n": 5,
1887
+ "rate": 0.4,
1888
+ "ci": [
1889
+ 0.1176182311592533,
1890
+ 0.769280067791163
1891
+ ]
1892
+ },
1893
+ "14": {
1894
+ "solved": 0,
1895
+ "n": 2,
1896
+ "rate": 0.0,
1897
+ "ci": [
1898
+ 0.0,
1899
+ 0.6576280471103807
1900
+ ]
1901
+ }
1902
+ },
1903
+ "sample_acc": 0.44511507220216606,
1904
+ "max_pruned": 15,
1905
+ "sampling": {
1906
+ "samples": 70912,
1907
+ "secs": 344.34613609313965,
1908
+ "gate_secs": 115.80984687805176,
1909
+ "samples_per_sec": 205.93232380809854
1910
+ }
1911
+ }
1912
+ }
1913
+ }