{
"methods": {
"original": {
"training_complete": true,
"evaluation_seconds": 3060.1390517000036,
"complete": true,
"test": {
"n": 2048,
"expected": 2048,
"complete": true,
"accuracy": 0.0029296875,
"interval": [
0.0009765625,
0.00537109375
],
"validity": 0.29296875,
"reward": 0.03193359375,
"generated_tokens": 95929,
"truncated": 529,
"correct_expression_structures": 6
},
"hard": {
"n": 1024,
"expected": 1024,
"complete": true,
"accuracy": 0.0009765625,
"interval": [
0.0,
0.0029296875
],
"validity": 0.1064453125,
"reward": 0.0115234375,
"generated_tokens": 59320,
"truncated": 333,
"correct_expression_structures": 1
},
"probes": {
"n": 16,
"expected": 16,
"complete": true,
"accuracy": 0.0,
"interval": [
0.0,
0.0
],
"validity": 0.25,
"reward": 0.025,
"generated_tokens": 823,
"truncated": 5,
"correct_expression_structures": 0
},
"sampled": {
"complete_problems": 256,
"expected": 256,
"generated_tokens": 229049,
"partial_problems": 0,
"pass1": 0.001220703125,
"pass8": 0.008854166666666666,
"mean_unique_correct_structures": 0.015625
},
"training": {},
"training_observations": {
"responses": 0,
"generated_tokens": 0,
"truncated": 0,
"weight_diagnostic": {
"generations_observed": 0,
"candidates_observed": 0,
"candidates_positive_novelty": 0,
"generations_positive_novelty": 0,
"generations_weight_changed": 0,
"candidates_weight_changed": 0,
"largest_absolute_weight_change": 0.0
}
},
"training_seconds": 0.0
},
"main": {
"training_complete": true,
"evaluation_seconds": 1940.7667252999672,
"complete": true,
"test": {
"n": 2048,
"expected": 2048,
"complete": true,
"accuracy": 0.01123046875,
"interval": [
0.0068359375,
0.01611328125
],
"validity": 0.76123046875,
"reward": 0.08623046875,
"generated_tokens": 41525,
"truncated": 0,
"correct_expression_structures": 23
},
"hard": {
"n": 1024,
"expected": 1024,
"complete": true,
"accuracy": 0.0,
"interval": [
0.0,
0.0
],
"validity": 0.0205078125,
"reward": 0.00205078125,
"generated_tokens": 23091,
"truncated": 0,
"correct_expression_structures": 0
},
"probes": {
"n": 16,
"expected": 16,
"complete": true,
"accuracy": 0.0,
"interval": [
0.0,
0.0
],
"validity": 0.875,
"reward": 0.08750000000000001,
"generated_tokens": 306,
"truncated": 0,
"correct_expression_structures": 0
},
"sampled": {
"complete_problems": 256,
"expected": 256,
"generated_tokens": 152280,
"partial_problems": 0,
"pass1": 0.002685546875,
"pass8": 0.020572916666666666,
"mean_unique_correct_structures": 0.0390625
},
"training": {
"generation": 19,
"generations": 19,
"complete": true,
"generated_tokens": 798675,
"responses": 27360
},
"training_observations": {
"responses": 27360,
"generated_tokens": 798675,
"truncated": 1999,
"weight_diagnostic": {
"generations_observed": 19,
"candidates_observed": 570,
"candidates_positive_novelty": 15,
"generations_positive_novelty": 10,
"generations_weight_changed": 9,
"candidates_weight_changed": 14,
"largest_absolute_weight_change": 0.0625
}
},
"training_seconds": 6078.265563964844
},
"reward_only": {
"training_complete": true,
"evaluation_seconds": 1949.0018551001413,
"complete": true,
"test": {
"n": 2048,
"expected": 2048,
"complete": true,
"accuracy": 0.01123046875,
"interval": [
0.0068359375,
0.01611328125
],
"validity": 0.7978515625,
"reward": 0.08989257812500001,
"generated_tokens": 42464,
"truncated": 0,
"correct_expression_structures": 23
},
"hard": {
"n": 1024,
"expected": 1024,
"complete": true,
"accuracy": 0.0009765625,
"interval": [
0.0,
0.0029296875
],
"validity": 0.03515625,
"reward": 0.00439453125,
"generated_tokens": 23582,
"truncated": 0,
"correct_expression_structures": 1
},
"probes": {
"n": 16,
"expected": 16,
"complete": true,
"accuracy": 0.0,
"interval": [
0.0,
0.0
],
"validity": 0.8125,
"reward": 0.08125000000000002,
"generated_tokens": 316,
"truncated": 0,
"correct_expression_structures": 0
},
"sampled": {
"complete_problems": 256,
"expected": 256,
"generated_tokens": 163148,
"partial_problems": 0,
"pass1": 0.002197265625,
"pass8": 0.016666666666666666,
"mean_unique_correct_structures": 0.03125
},
"training": {
"generation": 19,
"generations": 19,
"complete": true,
"generated_tokens": 814802,
"responses": 27360
},
"training_observations": {
"responses": 27360,
"generated_tokens": 814802,
"truncated": 2093,
"weight_diagnostic": {
"generations_observed": 19,
"candidates_observed": 570,
"candidates_positive_novelty": 13,
"generations_positive_novelty": 9,
"generations_weight_changed": 0,
"candidates_weight_changed": 0,
"largest_absolute_weight_change": 0.0
}
},
"training_seconds": 5984.517926454544
}
},
"examples": {
"original": [
{
"problem_id": "test-00000",
"prompt": "Use the numbers [49, 11, 13] exactly once each to make 73. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "49 * (11 - 13)",
"numbers": [
49,
11,
13
],
"target": 73,
"generated_tokens": 17,
"token_ids": [
27,
9217,
29,
19,
24,
353,
320,
16,
16,
481,
220,
16,
18,
12533,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "49 * (11 - 13)",
"structure": "(* (- 11 13) 49)",
"sample": null
},
{
"problem_id": "test-00001",
"prompt": "Use the numbers [21, 3, 42, 2] exactly once each to make 68. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "2 * (3 - 2) + 42 = 68",
"numbers": [
21,
3,
42,
2
],
"target": 68,
"generated_tokens": 23,
"token_ids": [
27,
9217,
29,
17,
353,
320,
18,
481,
220,
17,
8,
488,
220,
19,
17,
284,
220,
21,
23,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00002",
"prompt": "Use the numbers [28, 47, 11] exactly once each to make 355. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(28 * (47 - 11)) / 3",
"numbers": [
28,
47,
11
],
"target": 355,
"generated_tokens": 16,
"token_ids": [
7,
17,
23,
353,
320,
19,
22,
481,
220,
16,
16,
593,
608,
220,
18,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00003",
"prompt": "Use the numbers [48, 34, 24, 30] exactly once each to make 40. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(30 * (48 - 34)) / 24",
"numbers": [
48,
34,
24,
30
],
"target": 40,
"generated_tokens": 17,
"token_ids": [
7,
18,
15,
353,
320,
19,
23,
481,
220,
18,
19,
593,
608,
220,
17,
19,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(30 * (48 - 34)) / 24",
"structure": "(/ (* (- 48 34) 30) 24)",
"sample": null
},
{
"problem_id": "test-00004",
"prompt": "Use the numbers [45, 15, 23] exactly once each to make 83. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(23 * (15 - 45)) / 15",
"numbers": [
45,
15,
23
],
"target": 83,
"generated_tokens": 17,
"token_ids": [
7,
17,
18,
353,
320,
16,
20,
481,
220,
19,
20,
593,
608,
220,
16,
20,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00005",
"prompt": "Use the numbers [32, 10, 36, 40] exactly once each to make 360. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(36 * (10 - 32)) / 40",
"numbers": [
32,
10,
36,
40
],
"target": 360,
"generated_tokens": 17,
"token_ids": [
7,
18,
21,
353,
320,
16,
15,
481,
220,
18,
17,
593,
608,
220,
19,
15,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(36 * (10 - 32)) / 40",
"structure": "(/ (* (- 10 32) 36) 40)",
"sample": null
},
{
"problem_id": "test-00006",
"prompt": "Use the numbers [33, 6, 43] exactly once each to make 82. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "82 = (33 * 43 - 6) / 1",
"numbers": [
33,
6,
43
],
"target": 82,
"generated_tokens": 24,
"token_ids": [
27,
9217,
29,
23,
17,
284,
320,
18,
18,
353,
220,
19,
18,
481,
220,
21,
8,
608,
220,
16,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00007",
"prompt": "Use the numbers [40, 9, 23, 18] exactly once each to make 463. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "To solve this problem, we need to find a combination of the given numbers (40, 9, 23, 18) that equals 463 using only basic arithmetic operations (+, -, *, /). Let's explore different combinations:\n\n1. **Using Parentheses:**\n - Try \\( \\frac{40}{\\frac{9}{\\frac{23}{18}}} = 463 \\)\n - This is because:\n \\[\n \\frac{40}{\\frac{9}{\\frac{23}{18}}} = \\frac{40",
"numbers": [
40,
9,
23,
18
],
"target": 463,
"generated_tokens": 128,
"token_ids": [
1249,
11625,
419,
3491,
11,
582,
1184,
311,
1477,
264,
10601,
315,
279,
2661,
5109,
320,
19,
15,
11,
220,
24,
11,
220,
17,
18,
11,
220,
16,
23,
8,
429,
16819,
220,
19,
21,
18,
1667,
1172,
6770,
34784,
7525,
17973,
11,
85922,
11777,
608,
568,
6771,
594,
13186,
2155,
27459,
1447,
16,
13,
3070,
16429,
45686,
38322,
25,
1019,
256,
481,
9735,
17767,
1124,
37018,
90,
19,
15,
15170,
59,
37018,
90,
24,
15170,
59,
37018,
90,
17,
18,
15170,
16,
23,
75542,
284,
220,
19,
21,
18,
1124,
340,
256,
481,
1096,
374,
1576,
510,
257,
1124,
9640,
257,
1124,
37018,
90,
19,
15,
15170,
59,
37018,
90,
24,
15170,
59,
37018,
90,
17,
18,
15170,
16,
23,
75542,
284,
1124,
37018,
90,
19,
15
],
"truncated": true,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00008",
"prompt": "Use the numbers [18, 32, 37] exactly once each to make 698. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "18 * (32 - 37) = 698",
"numbers": [
18,
32,
37
],
"target": 698,
"generated_tokens": 23,
"token_ids": [
27,
9217,
29,
16,
23,
353,
320,
18,
17,
481,
220,
18,
22,
8,
284,
220,
21,
24,
23,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00009",
"prompt": "Use the numbers [6, 22, 24, 50] exactly once each to make 10. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(22 * (50 - 6) + 24) / 2",
"numbers": [
6,
22,
24,
50
],
"target": 10,
"generated_tokens": 20,
"token_ids": [
7,
17,
17,
353,
320,
20,
15,
481,
220,
21,
8,
488,
220,
17,
19,
8,
608,
220,
17,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00010",
"prompt": "Use the numbers [6, 39, 21] exactly once each to make 54. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "6 * (39 - 21)",
"numbers": [
6,
39,
21
],
"target": 54,
"generated_tokens": 16,
"token_ids": [
27,
9217,
29,
21,
353,
320,
18,
24,
481,
220,
17,
16,
12533,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "6 * (39 - 21)",
"structure": "(* (- 39 21) 6)",
"sample": null
},
{
"problem_id": "test-00011",
"prompt": "Use the numbers [1, 37, 20, 25] exactly once each to make 252. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "To solve the problem of making 252 using the numbers 1, 37, 20, and 25 exactly once each, we need to carefully consider how to combine these numbers using arithmetic operations (addition, subtraction, multiplication, division) and parentheses.\n\nLet's start by trying different combinations:\n\n1. **Using addition**: \n - \\( 25 \\times 37 = 925 \\)\n - Subtracting 20 from this result: \\( 925 - 20 = 905 \\)\n - Adding 25 to this result:",
"numbers": [
1,
37,
20,
25
],
"target": 252,
"generated_tokens": 128,
"token_ids": [
1249,
11625,
279,
3491,
315,
3259,
220,
17,
20,
17,
1667,
279,
5109,
220,
16,
11,
220,
18,
22,
11,
220,
17,
15,
11,
323,
220,
17,
20,
6896,
3055,
1817,
11,
582,
1184,
311,
15516,
2908,
1246,
311,
15963,
1493,
5109,
1667,
34784,
7525,
320,
718,
680,
11,
75240,
11,
46444,
11,
12804,
8,
323,
73975,
382,
10061,
594,
1191,
553,
4460,
2155,
27459,
1447,
16,
13,
3070,
16429,
5256,
95518,
715,
256,
481,
17767,
220,
17,
20,
1124,
15136,
220,
18,
22,
284,
220,
24,
17,
20,
1124,
340,
256,
481,
93210,
287,
220,
17,
15,
504,
419,
1102,
25,
17767,
220,
24,
17,
20,
481,
220,
17,
15,
284,
220,
24,
15,
20,
1124,
340,
256,
481,
30370,
220,
17,
20,
311,
419,
1102,
25
],
"truncated": true,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00012",
"prompt": "Use the numbers [38, 17, 37] exactly once each to make 16. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(38 - 17) * 37",
"numbers": [
38,
17,
37
],
"target": 16,
"generated_tokens": 13,
"token_ids": [
7,
18,
23,
481,
220,
16,
22,
8,
353,
220,
18,
22,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(38 - 17) * 37",
"structure": "(* (- 38 17) 37)",
"sample": null
},
{
"problem_id": "test-00013",
"prompt": "Use the numbers [38, 10, 23, 38] exactly once each to make 33. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(38 - (10 * 23)) / 38",
"numbers": [
38,
10,
23,
38
],
"target": 33,
"generated_tokens": 17,
"token_ids": [
7,
18,
23,
481,
320,
16,
15,
353,
220,
17,
18,
593,
608,
220,
18,
23,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(38 - (10 * 23)) / 38",
"structure": "(/ (- 38 (* 10 23)) 38)",
"sample": null
},
{
"problem_id": "test-00014",
"prompt": "Use the numbers [6, 35, 13] exactly once each to make 245. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(13 * (6 - 35) / 13)",
"numbers": [
6,
35,
13
],
"target": 245,
"generated_tokens": 17,
"token_ids": [
7,
16,
18,
353,
320,
21,
481,
220,
18,
20,
8,
608,
220,
16,
18,
8,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00015",
"prompt": "Use the numbers [49, 22, 24, 45] exactly once each to make 92. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(24 * (45 - 22)) / 49",
"numbers": [
49,
22,
24,
45
],
"target": 92,
"generated_tokens": 17,
"token_ids": [
7,
17,
19,
353,
320,
19,
20,
481,
220,
17,
17,
593,
608,
220,
19,
24,
151645
],
"truncated": false,
"batch_seconds": 7.101204900001903,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(24 * (45 - 22)) / 49",
"structure": "(/ (* (- 45 22) 24) 49)",
"sample": null
}
],
"main": [
{
"problem_id": "test-00000",
"prompt": "Use the numbers [49, 11, 13] exactly once each to make 73. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(13 - 49) * 11\n",
"numbers": [
49,
11,
13
],
"target": 73,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
16,
18,
481,
220,
19,
24,
8,
353,
220,
16,
16,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(13 - 49) * 11",
"structure": "(* (- 13 49) 11)",
"sample": null
},
{
"problem_id": "test-00001",
"prompt": "Use the numbers [21, 3, 42, 2] exactly once each to make 68. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(2 * (3 - 21)) + 42\n",
"numbers": [
21,
3,
42,
2
],
"target": 68,
"generated_tokens": 22,
"token_ids": [
27,
9217,
397,
7,
17,
353,
320,
18,
481,
220,
17,
16,
593,
488,
220,
19,
17,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(2 * (3 - 21)) + 42",
"structure": "(+ (* (- 3 21) 2) 42)",
"sample": null
},
{
"problem_id": "test-00002",
"prompt": "Use the numbers [28, 47, 11] exactly once each to make 355. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(28 - (47 * 11))\n",
"numbers": [
28,
47,
11
],
"target": 355,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
17,
23,
481,
320,
19,
22,
353,
220,
16,
16,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(28 - (47 * 11))",
"structure": "(- 28 (* 11 47))",
"sample": null
},
{
"problem_id": "test-00003",
"prompt": "Use the numbers [48, 34, 24, 30] exactly once each to make 40. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(30 - 48) * 34 / 24",
"numbers": [
48,
34,
24,
30
],
"target": 40,
"generated_tokens": 17,
"token_ids": [
7,
18,
15,
481,
220,
19,
23,
8,
353,
220,
18,
19,
608,
220,
17,
19,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(30 - 48) * 34 / 24",
"structure": "(/ (* (- 30 48) 34) 24)",
"sample": null
},
{
"problem_id": "test-00004",
"prompt": "Use the numbers [45, 15, 23] exactly once each to make 83. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(45 - 15) * 23\n",
"numbers": [
45,
15,
23
],
"target": 83,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
19,
20,
481,
220,
16,
20,
8,
353,
220,
17,
18,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(45 - 15) * 23",
"structure": "(* (- 45 15) 23)",
"sample": null
},
{
"problem_id": "test-00005",
"prompt": "Use the numbers [32, 10, 36, 40] exactly once each to make 360. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(36 * (40 - 32) / 10)\n",
"numbers": [
32,
10,
36,
40
],
"target": 360,
"generated_tokens": 24,
"token_ids": [
27,
9217,
397,
7,
18,
21,
353,
320,
19,
15,
481,
220,
18,
17,
8,
608,
220,
16,
15,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(36 * (40 - 32) / 10)",
"structure": "(/ (* (- 40 32) 36) 10)",
"sample": null
},
{
"problem_id": "test-00006",
"prompt": "Use the numbers [33, 6, 43] exactly once each to make 82. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(33 - 6) * 43\n",
"numbers": [
33,
6,
43
],
"target": 82,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
18,
481,
220,
21,
8,
353,
220,
19,
18,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(33 - 6) * 43",
"structure": "(* (- 33 6) 43)",
"sample": null
},
{
"problem_id": "test-00007",
"prompt": "Use the numbers [40, 9, 23, 18] exactly once each to make 463. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(18 - 9) * 40 + 23\n",
"numbers": [
40,
9,
23,
18
],
"target": 463,
"generated_tokens": 23,
"token_ids": [
27,
9217,
397,
7,
16,
23,
481,
220,
24,
8,
353,
220,
19,
15,
488,
220,
17,
18,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(18 - 9) * 40 + 23",
"structure": "(+ (* (- 18 9) 40) 23)",
"sample": null
},
{
"problem_id": "test-00008",
"prompt": "Use the numbers [18, 32, 37] exactly once each to make 698. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(37 - (18 * 32))\n",
"numbers": [
18,
32,
37
],
"target": 698,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
22,
481,
320,
16,
23,
353,
220,
18,
17,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(37 - (18 * 32))",
"structure": "(- 37 (* 18 32))",
"sample": null
},
{
"problem_id": "test-00009",
"prompt": "Use the numbers [6, 22, 24, 50] exactly once each to make 10. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "(22 - (50 - 6)) * 24",
"numbers": [
6,
22,
24,
50
],
"target": 10,
"generated_tokens": 16,
"token_ids": [
7,
17,
17,
481,
320,
20,
15,
481,
220,
21,
593,
353,
220,
17,
19,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(22 - (50 - 6)) * 24",
"structure": "(* (- 22 (- 50 6)) 24)",
"sample": null
},
{
"problem_id": "test-00010",
"prompt": "Use the numbers [6, 39, 21] exactly once each to make 54. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(6 * (39 - 21))\n",
"numbers": [
6,
39,
21
],
"target": 54,
"generated_tokens": 18,
"token_ids": [
27,
9217,
397,
7,
21,
353,
320,
18,
24,
481,
220,
17,
16,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(6 * (39 - 21))",
"structure": "(* (- 39 21) 6)",
"sample": null
},
{
"problem_id": "test-00011",
"prompt": "Use the numbers [1, 37, 20, 25] exactly once each to make 252. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(25 - 20) * 1\n",
"numbers": [
1,
37,
20,
25
],
"target": 252,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
17,
20,
481,
220,
17,
15,
8,
353,
220,
16,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00012",
"prompt": "Use the numbers [38, 17, 37] exactly once each to make 16. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(38 - 17) * 37\n",
"numbers": [
38,
17,
37
],
"target": 16,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
18,
23,
481,
220,
16,
22,
8,
353,
220,
18,
22,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(38 - 17) * 37",
"structure": "(* (- 38 17) 37)",
"sample": null
},
{
"problem_id": "test-00013",
"prompt": "Use the numbers [38, 10, 23, 38] exactly once each to make 33. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(38 - (10 * 23))\n",
"numbers": [
38,
10,
23,
38
],
"target": 33,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
23,
481,
320,
16,
15,
353,
220,
17,
18,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00014",
"prompt": "Use the numbers [6, 35, 13] exactly once each to make 245. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(13 * (6 - 35) / 1)\n",
"numbers": [
6,
35,
13
],
"target": 245,
"generated_tokens": 22,
"token_ids": [
27,
9217,
397,
7,
16,
18,
353,
320,
21,
481,
220,
18,
20,
8,
608,
220,
16,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00015",
"prompt": "Use the numbers [49, 22, 24, 45] exactly once each to make 92. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(49 - (22 * 24) / 45)\n",
"numbers": [
49,
22,
24,
45
],
"target": 92,
"generated_tokens": 24,
"token_ids": [
27,
9217,
397,
7,
19,
24,
481,
320,
17,
17,
353,
220,
17,
19,
8,
608,
220,
19,
20,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.5798094000056153,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(49 - (22 * 24) / 45)",
"structure": "(- 49 (/ (* 22 24) 45))",
"sample": null
}
],
"reward_only": [
{
"problem_id": "test-00000",
"prompt": "Use the numbers [49, 11, 13] exactly once each to make 73. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(49 - 11) * (11 + 13)\n",
"numbers": [
49,
11,
13
],
"target": 73,
"generated_tokens": 24,
"token_ids": [
27,
9217,
397,
7,
19,
24,
481,
220,
16,
16,
8,
353,
320,
16,
16,
488,
220,
16,
18,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00001",
"prompt": "Use the numbers [21, 3, 42, 2] exactly once each to make 68. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(2 * (3 - 21)) + 42\n",
"numbers": [
21,
3,
42,
2
],
"target": 68,
"generated_tokens": 22,
"token_ids": [
27,
9217,
397,
7,
17,
353,
320,
18,
481,
220,
17,
16,
593,
488,
220,
19,
17,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(2 * (3 - 21)) + 42",
"structure": "(+ (* (- 3 21) 2) 42)",
"sample": null
},
{
"problem_id": "test-00002",
"prompt": "Use the numbers [28, 47, 11] exactly once each to make 355. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(28 - (47 * 11))\n",
"numbers": [
28,
47,
11
],
"target": 355,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
17,
23,
481,
320,
19,
22,
353,
220,
16,
16,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(28 - (47 * 11))",
"structure": "(- 28 (* 11 47))",
"sample": null
},
{
"problem_id": "test-00003",
"prompt": "Use the numbers [48, 34, 24, 30] exactly once each to make 40. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(30 - 34) * 48\n",
"numbers": [
48,
34,
24,
30
],
"target": 40,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
18,
15,
481,
220,
18,
19,
8,
353,
220,
19,
23,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00004",
"prompt": "Use the numbers [45, 15, 23] exactly once each to make 83. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(45 - 15) * 23\n",
"numbers": [
45,
15,
23
],
"target": 83,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
19,
20,
481,
220,
16,
20,
8,
353,
220,
17,
18,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(45 - 15) * 23",
"structure": "(* (- 45 15) 23)",
"sample": null
},
{
"problem_id": "test-00005",
"prompt": "Use the numbers [32, 10, 36, 40] exactly once each to make 360. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(32 * (10 - 36) / 40)\n",
"numbers": [
32,
10,
36,
40
],
"target": 360,
"generated_tokens": 24,
"token_ids": [
27,
9217,
397,
7,
18,
17,
353,
320,
16,
15,
481,
220,
18,
21,
8,
608,
220,
19,
15,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(32 * (10 - 36) / 40)",
"structure": "(/ (* (- 10 36) 32) 40)",
"sample": null
},
{
"problem_id": "test-00006",
"prompt": "Use the numbers [33, 6, 43] exactly once each to make 82. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(33 - 6) * 43\n",
"numbers": [
33,
6,
43
],
"target": 82,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
18,
481,
220,
21,
8,
353,
220,
19,
18,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(33 - 6) * 43",
"structure": "(* (- 33 6) 43)",
"sample": null
},
{
"problem_id": "test-00007",
"prompt": "Use the numbers [40, 9, 23, 18] exactly once each to make 463. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(40 - (9 + 23) * 18)\n",
"numbers": [
40,
9,
23,
18
],
"target": 463,
"generated_tokens": 23,
"token_ids": [
27,
9217,
397,
7,
19,
15,
481,
320,
24,
488,
220,
17,
18,
8,
353,
220,
16,
23,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(40 - (9 + 23) * 18)",
"structure": "(- 40 (* (+ 23 9) 18))",
"sample": null
},
{
"problem_id": "test-00008",
"prompt": "Use the numbers [18, 32, 37] exactly once each to make 698. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(37 - (18 * 32))\n",
"numbers": [
18,
32,
37
],
"target": 698,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
22,
481,
320,
16,
23,
353,
220,
18,
17,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(37 - (18 * 32))",
"structure": "(- 37 (* 18 32))",
"sample": null
},
{
"problem_id": "test-00009",
"prompt": "Use the numbers [6, 22, 24, 50] exactly once each to make 10. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(22 * (50 - 6) + 24)\n",
"numbers": [
6,
22,
24,
50
],
"target": 10,
"generated_tokens": 23,
"token_ids": [
27,
9217,
397,
7,
17,
17,
353,
320,
20,
15,
481,
220,
21,
8,
488,
220,
17,
19,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(22 * (50 - 6) + 24)",
"structure": "(+ (* (- 50 6) 22) 24)",
"sample": null
},
{
"problem_id": "test-00010",
"prompt": "Use the numbers [6, 39, 21] exactly once each to make 54. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(6 * (39 - 21))\n",
"numbers": [
6,
39,
21
],
"target": 54,
"generated_tokens": 18,
"token_ids": [
27,
9217,
397,
7,
21,
353,
320,
18,
24,
481,
220,
17,
16,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(6 * (39 - 21))",
"structure": "(* (- 39 21) 6)",
"sample": null
},
{
"problem_id": "test-00011",
"prompt": "Use the numbers [1, 37, 20, 25] exactly once each to make 252. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(25 * (1 - 37 / 20))\n",
"numbers": [
1,
37,
20,
25
],
"target": 252,
"generated_tokens": 22,
"token_ids": [
27,
9217,
397,
7,
17,
20,
353,
320,
16,
481,
220,
18,
22,
608,
220,
17,
15,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(25 * (1 - 37 / 20))",
"structure": "(* (- 1 (/ 37 20)) 25)",
"sample": null
},
{
"problem_id": "test-00012",
"prompt": "Use the numbers [38, 17, 37] exactly once each to make 16. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(38 - 17) * 37\n",
"numbers": [
38,
17,
37
],
"target": 16,
"generated_tokens": 20,
"token_ids": [
27,
9217,
397,
7,
18,
23,
481,
220,
16,
22,
8,
353,
220,
18,
22,
198,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(38 - 17) * 37",
"structure": "(* (- 38 17) 37)",
"sample": null
},
{
"problem_id": "test-00013",
"prompt": "Use the numbers [38, 10, 23, 38] exactly once each to make 33. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(38 - (10 * 23))\n",
"numbers": [
38,
10,
23,
38
],
"target": 33,
"generated_tokens": 19,
"token_ids": [
27,
9217,
397,
7,
18,
23,
481,
320,
16,
15,
353,
220,
17,
18,
1171,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00014",
"prompt": "Use the numbers [6, 35, 13] exactly once each to make 245. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(13 * (6 - 35) / 1)\n",
"numbers": [
6,
35,
13
],
"target": 245,
"generated_tokens": 22,
"token_ids": [
27,
9217,
397,
7,
16,
18,
353,
320,
21,
481,
220,
18,
20,
8,
608,
220,
16,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.0,
"correct": false,
"valid": false,
"expression": "",
"structure": "",
"sample": null
},
{
"problem_id": "test-00015",
"prompt": "Use the numbers [49, 22, 24, 45] exactly once each to make 92. You may use +, -, *, / and parentheses. Do not use other numbers or operations. Return only expression.",
"response": "\n(49 - (22 * 24) / 45)\n",
"numbers": [
49,
22,
24,
45
],
"target": 92,
"generated_tokens": 24,
"token_ids": [
27,
9217,
397,
7,
19,
24,
481,
320,
17,
17,
353,
220,
17,
19,
8,
608,
220,
19,
20,
340,
522,
9217,
29,
151645
],
"truncated": false,
"batch_seconds": 1.518910999999207,
"reward": 0.1,
"correct": false,
"valid": true,
"expression": "(49 - (22 * 24) / 45)",
"structure": "(- 49 (/ (* 22 24) 45))",
"sample": null
}
]
},
"primary": {
"difference": 0.0,
"low": -0.00244140625,
"high": 0.00244140625,
"n": 2048
},
"baseline_comparisons": {
"main": {
"difference": 0.00830078125,
"low": 0.004382324218750011,
"high": 0.0126953125,
"n": 2048
},
"reward_only": {
"difference": 0.00830078125,
"low": 0.00390625,
"high": 0.0126953125,
"n": 2048
}
},
"model_id": "Qwen/Qwen2.5-0.5B-Instruct",
"seed": 0,
"initializations": {
"main": {
"checked_at": "2026-11-07T04:57:19.449952+00:00",
"condition": "main",
"generation": 0,
"model_id": "Qwen/Qwen2.5-0.5B-Instruct",
"model_revision": "7ae557604adf67be50417f59c2c2f167def9a775",
"training_code_revision": "9f8d30ef0103e3532ebefeafe7f55b8f8d163a80",
"weighting": "zscore_task_plus_novelty_v1",
"novelty_weight": 1.0,
"unique_tensors": 290,
"parameter_elements": 494032768,
"architecture": {
"num_hidden_layers": 24,
"hidden_size": 896,
"intermediate_size": 4864,
"num_attention_heads": 14,
"num_key_value_heads": 2,
"vocab_size": 151936,
"max_position_embeddings": 32768,
"tie_word_embeddings": true
},
"all_pretrained_tensors_exactly_equal_in_fp32": true,
"initial_archive_empty": true,
"absolute_deadline": 1794067007.1504683
},
"reward_only": {
"checked_at": "2026-11-07T06:38:33.868772+00:00",
"condition": "reward_only",
"generation": 0,
"model_id": "Qwen/Qwen2.5-0.5B-Instruct",
"model_revision": "7ae557604adf67be50417f59c2c2f167def9a775",
"training_code_revision": "9f8d30ef0103e3532ebefeafe7f55b8f8d163a80",
"weighting": "zscore_task_plus_novelty_v1",
"novelty_weight": 1.0,
"unique_tensors": 290,
"parameter_elements": 494032768,
"architecture": {
"num_hidden_layers": 24,
"hidden_size": 896,
"intermediate_size": 4864,
"num_attention_heads": 14,
"num_key_value_heads": 2,
"vocab_size": 151936,
"max_position_embeddings": 32768,
"tie_word_embeddings": true
},
"all_pretrained_tensors_exactly_equal_in_fp32": true,
"initial_archive_empty": true,
"absolute_deadline": 1794067007.1504683
}
},
"novelty_weight": 1.0,
"finding": "The directional improvement hypothesis was not supported in this seed: main accuracy did not exceed reward-only ES.",
"generated_at": "2026-11-07T09:33:53.529406+00:00",
"interval_scope": "Problem uncertainty conditional on one training seed; excludes seed variability.",
"interval_caveat": "Percentile bootstrap intervals can collapse at observed boundaries or when paired outcomes are identical. A zero-width interval does not prove zero population error or equivalence on unseen problems.",
"failures": [
{
"time": 1794026144.5639172,
"phase": "calibration",
"exception": "PermissionError",
"message": "[WinError 5] Access is denied: 'artifacts\\\\pilot\\\\original\\\\evaluation.json.tmp' -> 'artifacts\\\\pilot\\\\original\\\\evaluation.json'"
}
],
"protocol_amendments": [
{
"id": "A",
"applied_at": 1794027075.1466103,
"phase_before": "calibration",
"registration_revision": "158becd",
"implementation_revision": "9f8d30ef0103e3532ebefeafe7f55b8f8d163a80",
"reason": "User-requested source alignment before either trained condition; ExTra post-normalization addition and ES at Scale population z-scores",
"initialization": "Both arms start from original pretrained weights; no prior training checkpoints exist",
"original_model_evaluation_already_observed": true,
"changes": {
"novelty_weight": {
"before": 0.25,
"after": 1.0
},
"weighting": {
"before": null,
"after": "zscore_task_plus_novelty_v1"
}
},
"absolute_deadline": 1794067007.1504683
}
],
"execution_repairs": [
{
"time": 1794026373.4031878,
"revision": "4cbc4db74b77570c4b66793a544e7f02dda0287b",
"change": "Bounded retry for Windows atomic replacement locks; no numerical training changes",
"resume_from": "last committed original-model evaluation batch",
"absolute_deadline": 1794067007.1504683
}
],
"recovery_audit": {
"checked_at": "2026-11-07T05:41:15.301+00:00",
"condition": "original",
"split": "sampled",
"before_sha256": "10f4d15e6946fa25b09bf78a635003b3a6f4de637a26f4b05769485a20f52af2",
"before_rows": 2128,
"after_rows_at_check": 3136,
"replayed_batch_start": 2112,
"replayed_batch_rows": 16,
"all_prior_and_replayed_generation_fields_equal": true,
"excluded_field": "batch_seconds",
"absolute_deadline": 1794067007.1504683,
"scope": "Observed storage-failure recovery; not a claim about arbitrary restarts or environments"
},
"logged_generation_work": [
{
"condition": "calibration",
"phase": "calibration",
"responses": 144,
"generated_tokens": 4602
},
{
"condition": "calibration",
"phase": "calibration_validation",
"responses": 64,
"generated_tokens": 3018
},
{
"condition": "calibration",
"phase": "calibration_warmup",
"responses": 16,
"generated_tokens": 985
},
{
"condition": "main",
"phase": "evaluation",
"responses": 7184,
"generated_tokens": 217202
},
{
"condition": "main",
"phase": "train",
"responses": 27360,
"generated_tokens": 798675
},
{
"condition": "main",
"phase": "validation",
"responses": 256,
"generated_tokens": 5151
},
{
"condition": "original",
"phase": "evaluation",
"responses": 7200,
"generated_tokens": 385831
},
{
"condition": "reward_only",
"phase": "evaluation",
"responses": 7184,
"generated_tokens": 229510
},
{
"condition": "reward_only",
"phase": "train",
"responses": 27360,
"generated_tokens": 814802
},
{
"condition": "reward_only",
"phase": "validation",
"responses": 256,
"generated_tokens": 5142
}
],
"elapsed_seconds": 20151.339267730713
}