Asilarkness commited on
Commit
d9b05d6
·
verified ·
1 Parent(s): bc51733

Add v3 evaluation artifact V3_FINAL_DECISION.json

Browse files
candidates/budgie-alignment-v2/reasoning-sft-then-rl-v3/recovery/V3_FINAL_DECISION.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "date": "2026-08-26",
3
+ "stage": "reasoning SFT then RL v3",
4
+ "status": "Stage A rejected; Stage B RL not started; leader unchanged",
5
+ "base": "verified-math-a025",
6
+ "data": {
7
+ "scanned": 5251,
8
+ "verified_pool": 1600,
9
+ "train": 800,
10
+ "dev": 208,
11
+ "difficulty_train": {
12
+ "medium": 511,
13
+ "hard": 289
14
+ },
15
+ "sources_train": {
16
+ "olympiads": 357,
17
+ "aops_forum": 225,
18
+ "cn_contest": 154,
19
+ "amc_aime": 32,
20
+ "inequalities": 19,
21
+ "olympiads_ref": 10,
22
+ "number_theory": 3
23
+ },
24
+ "problem_types_train": {
25
+ "Number Theory": 179,
26
+ "Algebra": 228,
27
+ "Geometry": 193,
28
+ "Combinatorics": 119,
29
+ "Calculus": 19,
30
+ "Logic and Puzzles": 29,
31
+ "Other": 11,
32
+ "Inequalities": 22
33
+ },
34
+ "rejections": {
35
+ "verify": 2910,
36
+ "length": 740
37
+ },
38
+ "benchmark_family_rows": 0,
39
+ "decontamination": "normalized exact + shared fixed-gate 12-token windows + source bans",
40
+ "verification": "math_verify(reference solution, answer)",
41
+ "seconds": 69.16955494880676,
42
+ "difficulty_dev": {
43
+ "hard": 78,
44
+ "medium": 130
45
+ }
46
+ },
47
+ "encoding": {
48
+ "raw_records": 926,
49
+ "encoded": 908,
50
+ "math": 800,
51
+ "info": 108,
52
+ "mean_tokens": 576.6332599118942,
53
+ "mean_response_tokens": 441.8458149779736,
54
+ "max_tokens": 1536,
55
+ "seconds": 14.342991828918457
56
+ },
57
+ "training": {
58
+ "method": "one-epoch long verified reasoning SFT with sampled-action leader KL",
59
+ "base": "verified-math-a025",
60
+ "encoded_rows": 908,
61
+ "math_rows": 800,
62
+ "info_rows": 108,
63
+ "updates": 227,
64
+ "grad_accum": 4,
65
+ "peak_lr": 3e-07,
66
+ "kl_beta": 0.12,
67
+ "max_length": 1536,
68
+ "source_path": "/marimo/budgie_information_rl_v1/v3_reasoning_sft_source.pt",
69
+ "parameter_count_change": 0,
70
+ "benchmark_family_rows": 0
71
+ },
72
+ "results": {
73
+ "information_dev": {
74
+ "leader": "3/21, loops 0",
75
+ "alpha_0_005": "4/21, loops 0",
76
+ "smaller_alphas": "3/21"
77
+ },
78
+ "random_verified_math_dev_32": {
79
+ "leader": "1/32",
80
+ "alpha_0_005": "1/32"
81
+ },
82
+ "fixed_reasoning": {
83
+ "leader": {
84
+ "gsm": "5/30",
85
+ "math": "3/15",
86
+ "arc": "11/30",
87
+ "folio": "13/30"
88
+ },
89
+ "alpha_0_005": {
90
+ "gsm": "4/30",
91
+ "math": "2/15",
92
+ "arc": "11/30",
93
+ "folio": "13/30"
94
+ }
95
+ }
96
+ },
97
+ "decision": "Reject reasoning SFT Stage A because the only alpha with an information-management gain regresses fixed GSM and MATH. Do not run follow-on RL. Smaller alphas have no held-out information gain. Keep verified-math-a025.",
98
+ "parameter_count_change": 0,
99
+ "benchmark_family_rows": 0,
100
+ "public_root_changed": false
101
+ }