Label-Free Bayesian Truth Serum Reward / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "belief_sensitive_peer_prediction",
4 "model": "mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.003
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.016249999636784196
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.01499999943189323
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.019999999552965164
29 }
30 ],
31 "full": {
32 "mean": 0.016874999273568392,
33 "std": 0.007473912417713013,
34 "per_seed": [
35 0.009999999776482582,
36 0.019999999552965164,
37 0.02499999850988388,
38 0.004999999888241291,
39 0.014999999664723873,
40 0.02499999850988388,
41 0.009999999776482582,
42 0.02499999850988388
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.3287499975413084,
49 "std": 0.18487749114388957,
50 "per_seed": [
51 0.10999999940395355,
52 0.16500000655651093,
53 0.6800000071525574,
54 0.25999999046325684,
55 0.17000000178813934,
56 0.3050000071525574,
57 0.5199999809265137,
58 0.41999998688697815
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": 0.31187499826774,
64 "idea_wins": 0,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 0.09999999962747097,
68 0.14500000700354576,
69 0.6550000086426735,
70 0.25499999057501554,
71 0.15500000212341547,
72 0.2800000086426735,
73 0.5099999811500311,
74 0.39499998837709427
75 ],
76 "p_value": 0.0081,
77 "mde": 0.16284215222207377,
78 "mde_rel_pct": 964.9905732270832,
79 "verdict": "idea worse (significant)",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "custom_track": {
84 "name": "belief_sensitive_peer_prediction",
85 "file": "bts_track.py",
86 "domain": "rlhf"
87 },
88 "idea_sweep": [
89 {
90 "cfg": {
91 "lr": 0.001
92 },
93 "result": {
94 "mean": 0.4612500015646219,
95 "std": 0.25068343891180656,
96 "per_seed": [
97 0.375,
98 0.675000011920929,
99 0.8999999761581421,
100 0.5400000214576721,
101 0.4050000011920929,
102 0.17499999701976776,
103 0.5550000071525574,
104 0.06499999761581421
105 ],
106 "n": 8
107 }
108 },
109 {
110 "cfg": {
111 "lr": 0.003
112 },
113 "result": {
114 "mean": 0.3287499975413084,
115 "std": 0.18487749114388957,
116 "per_seed": [
117 0.10999999940395355,
118 0.16500000655651093,
119 0.6800000071525574,
120 0.25999999046325684,
121 0.17000000178813934,
122 0.3050000071525574,
123 0.5199999809265137,
124 0.41999998688697815
125 ],
126 "n": 8
127 }
128 },
129 {
130 "cfg": {
131 "lr": 0.01
132 },
133 "result": {
134 "mean": 0.5187499970197678,
135 "std": 0.04385701432651679,
136 "per_seed": [
137 0.5149999856948853,
138 0.45500001311302185,
139 0.5099999904632568,
140 0.5299999713897705,
141 0.6150000095367432,
142 0.47999998927116394,
143 0.5099999904632568,
144 0.5350000262260437
145 ],
146 "n": 8
147 }
148 }
149 ],
150 "mechanism_signature": {
151 "group_size": 8,
152 "n_model_completions": 1024,
153 "predicted_mean_frequency": 0.5131519436836243,
154 "observed_mean_frequency": 0.5771484375,
155 "predicted_observed_correlation": 0.400832738120897,
156 "log_ratio_max_error": 0.0,
157 "confirmed": true
158 }
159 }
160}