logicless commited on
Commit
52ff3b9
·
verified ·
1 Parent(s): 4914175

Add final evaluation and training results

Browse files
results/final_comparison.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "count": 240,
3
+ "evaluation": "Frozen 240-prompt LatentMD subset scored with local_structural_contract_v1",
4
+ "runs": {
5
+ "base": {
6
+ "full_contract_pass_rate": 0.08333333333333333,
7
+ "passed": 20
8
+ },
9
+ "run1_broad_sft": {
10
+ "full_contract_pass_rate": 0.32083333333333336,
11
+ "passed": 77
12
+ },
13
+ "run2_wrapper_booster": {
14
+ "full_contract_pass_rate": 0.825,
15
+ "passed": 198
16
+ }
17
+ },
18
+ "run2_axis_results": {
19
+ "A1": {
20
+ "count": 80,
21
+ "full_contract_pass_rate": 0.825,
22
+ "passed": 66
23
+ },
24
+ "A2": {
25
+ "count": 80,
26
+ "full_contract_pass_rate": 0.7875,
27
+ "passed": 63
28
+ },
29
+ "A3": {
30
+ "count": 80,
31
+ "full_contract_pass_rate": 0.8625,
32
+ "passed": 69
33
+ }
34
+ },
35
+ "run2_check_pass_rates": {
36
+ "balanced_fences": 0.975,
37
+ "blockquote_with_citation": 0.95,
38
+ "language_code_examples": 0.8791666666666667,
39
+ "numbered_list": 0.9666666666666667,
40
+ "outer_wrapper": 1.0,
41
+ "parse_success": 1.0,
42
+ "raw_markdown_source": 0.9333333333333333,
43
+ "table": 0.9833333333333333
44
+ },
45
+ "improvement": {
46
+ "base_to_run2_absolute": 0.7416666666666667,
47
+ "base_to_run2_relative_multiplier": 9.9,
48
+ "run1_to_run2_absolute": 0.5041666666666667,
49
+ "run1_to_run2_relative_multiplier": 2.571428571428571
50
+ },
51
+ "interpretation": "Targeted short contrastive SFT fixed the outer-wrapper decision without causing wrapper-policy regressions on A1 or A3. Remaining failures are primarily inner B3/B4 structure and occasional unbalanced generation."
52
+ }
results/wrapper_run2_latentmd_metrics.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "by_axis_cell": {
3
+ "A1_B1": {
4
+ "count": 20,
5
+ "pass_rate": 1.0,
6
+ "passed": 20
7
+ },
8
+ "A1_B2": {
9
+ "count": 20,
10
+ "pass_rate": 0.9,
11
+ "passed": 18
12
+ },
13
+ "A1_B3": {
14
+ "count": 20,
15
+ "pass_rate": 0.7,
16
+ "passed": 14
17
+ },
18
+ "A1_B4": {
19
+ "count": 20,
20
+ "pass_rate": 0.7,
21
+ "passed": 14
22
+ },
23
+ "A2_B1": {
24
+ "count": 20,
25
+ "pass_rate": 0.85,
26
+ "passed": 17
27
+ },
28
+ "A2_B2": {
29
+ "count": 20,
30
+ "pass_rate": 0.95,
31
+ "passed": 19
32
+ },
33
+ "A2_B3": {
34
+ "count": 20,
35
+ "pass_rate": 0.85,
36
+ "passed": 17
37
+ },
38
+ "A2_B4": {
39
+ "count": 20,
40
+ "pass_rate": 0.5,
41
+ "passed": 10
42
+ },
43
+ "A3_B1": {
44
+ "count": 20,
45
+ "pass_rate": 1.0,
46
+ "passed": 20
47
+ },
48
+ "A3_B2": {
49
+ "count": 20,
50
+ "pass_rate": 0.85,
51
+ "passed": 17
52
+ },
53
+ "A3_B3": {
54
+ "count": 20,
55
+ "pass_rate": 0.85,
56
+ "passed": 17
57
+ },
58
+ "A3_B4": {
59
+ "count": 20,
60
+ "pass_rate": 0.75,
61
+ "passed": 15
62
+ }
63
+ },
64
+ "check_pass_rates": {
65
+ "balanced_fences": 0.975,
66
+ "blockquote_with_citation": 0.95,
67
+ "language_code_examples": 0.8791666666666667,
68
+ "numbered_list": 0.9666666666666667,
69
+ "outer_wrapper": 1.0,
70
+ "parse_success": 1.0,
71
+ "raw_markdown_source": 0.9333333333333333,
72
+ "table": 0.9833333333333333
73
+ },
74
+ "count": 240,
75
+ "failure_counts": {
76
+ "balanced_fences": 6,
77
+ "blockquote_with_citation": 12,
78
+ "language_code_examples": 29,
79
+ "numbered_list": 8,
80
+ "raw_markdown_source": 16,
81
+ "table": 4
82
+ },
83
+ "full_structural_contract_pass_rate": 0.825,
84
+ "scoring_method": {
85
+ "description": "Deterministic parser- and prompt-contract-based scoring written for this experiment. It does not assess arbitrary task-answer semantic correctness and is not LatentMD's authors' official evaluator.",
86
+ "name": "local_structural_contract_v1"
87
+ }
88
+ }
results/wrapper_run2_training_summary.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adapter": "adapters/qwen25-coder-3b-markdown-reliability-wrapper-run2",
3
+ "adapter_sha256": "a892954dbf11822d6f4ab1ef50328b6ade8ca3fad072b8706eb319dfaa27e5d5",
4
+ "adapter_size_bytes": 119789457,
5
+ "base_adapter": "adapters/qwen25-coder-3b-markdown-reliability-lora/adapters.safetensors",
6
+ "batch_size": 2,
7
+ "effective_batch_size": 8,
8
+ "final_iteration": 300,
9
+ "final_reported_train_loss": 0.176,
10
+ "final_validation_loss": 0.266,
11
+ "grad_accumulation_steps": 4,
12
+ "learning_rate": 2e-06,
13
+ "optimizer_updates": 75,
14
+ "trained_tokens": 91031,
15
+ "training_example_presentations": 600,
16
+ "validation_loss_curve": {
17
+ "1": 1.334,
18
+ "50": 0.526,
19
+ "100": 0.406,
20
+ "150": 0.326,
21
+ "200": 0.288,
22
+ "250": 0.273,
23
+ "300": 0.266
24
+ }
25
+ }