supermix-v74 / results /problem_solving_n500.json
Kai9987kai's picture
Supermix v74: 8.6M-parameter arithmetic problem solver (0.894 on n=500)
718b458 verified
Raw
History Blame Contribute Delete
6 kB
{
"schema": "supermix-v65-problem-solving-accuracy-v1",
"checkpoint": "output\\v74_broad\\v74_broad.pt",
"dev_loss": 0.065107,
"problems": 500,
"unparsed_replies": 0,
"by_source": {
"novel": {
"n": 500,
"correct": 447,
"tasks": {
"algebra_one_step": {
"n": 56,
"correct": 50,
"accuracy": 0.8929,
"median_relative_error": 0.0,
"within_10_percent": 0.8929
},
"arithmetic": {
"n": 56,
"correct": 50,
"accuracy": 0.8929,
"median_relative_error": 0.0,
"within_10_percent": 1.0
},
"average": {
"n": 56,
"correct": 33,
"accuracy": 0.5893,
"median_relative_error": 0.0,
"within_10_percent": 0.9821
},
"division": {
"n": 55,
"correct": 55,
"accuracy": 1.0,
"median_relative_error": 0.0,
"within_10_percent": 1.0
},
"multiplication": {
"n": 55,
"correct": 55,
"accuracy": 1.0,
"median_relative_error": 0.0,
"within_10_percent": 1.0
},
"percent": {
"n": 56,
"correct": 42,
"accuracy": 0.75,
"median_relative_error": 0.0,
"within_10_percent": 0.7679
},
"sequence": {
"n": 55,
"correct": 54,
"accuracy": 0.9818,
"median_relative_error": 0.0,
"within_10_percent": 0.9818
},
"two_step": {
"n": 55,
"correct": 54,
"accuracy": 0.9818,
"median_relative_error": 0.0,
"within_10_percent": 0.9818
},
"word_problem": {
"n": 56,
"correct": 54,
"accuracy": 0.9643,
"median_relative_error": 0.0,
"within_10_percent": 0.9643
}
},
"accuracy": 0.894,
"median_relative_error": 0.0,
"within_10_percent": 0.952
}
},
"memorisation_gap": null,
"examples": [
{
"source": "novel",
"task": "arithmetic",
"prompt": "Solve this basic math problem: 524 - 305",
"reply": "500 - 300 = 200, 24 - 5 = 19, total 219",
"expected": 219.0,
"predicted": 219.0,
"correct": true
},
{
"source": "novel",
"task": "percent",
"prompt": "What is 20% of 1749?",
"reply": "1 percent of 1749 = 17.49, times 20 = 349.8, total 349.8",
"expected": 349.8,
"predicted": 349.8,
"correct": true
},
{
"source": "novel",
"task": "average",
"prompt": "Find the average (mean) of these numbers: 61, 63, 72, 61",
"reply": "sum: 61 then 124 then 196 then 257, total 257, divide by 4, total 64.25",
"expected": 64.25,
"predicted": 64.25,
"correct": true
},
{
"source": "novel",
"task": "algebra_one_step",
"prompt": "Solve for x: x + 29 = 34",
"reply": "subtract 29 from both sides, 34 - 29 = 5, total 5",
"expected": 5.0,
"predicted": 5.0,
"correct": true
},
{
"source": "novel",
"task": "word_problem",
"prompt": "A student has 63 stickers. They get 57 more and then give away 45. How many stickers do th",
"reply": "63 + 57 = 120, 120 - 45 = 75, total 75",
"expected": 75.0,
"predicted": 75.0,
"correct": true
},
{
"source": "novel",
"task": "multiplication",
"prompt": "Solve this basic math problem: 39 x 3",
"reply": "30 x 3 = 90, 9 x 3 = 27, total 117",
"expected": 117.0,
"predicted": 117.0,
"correct": true
},
{
"source": "novel",
"task": "division",
"prompt": "Solve this basic math problem: 243 / 9",
"reply": "180 / 9 = 20, 63 / 9 = 7, total 27",
"expected": 27.0,
"predicted": 27.0,
"correct": true
},
{
"source": "novel",
"task": "sequence",
"prompt": "What comes next in the sequence: 15, 27, 39, 51?",
"reply": "27 - 15 = 12, difference 12, 51 + 12 = 63, total 63",
"expected": 63.0,
"predicted": 63.0,
"correct": true
},
{
"source": "novel",
"task": "two_step",
"prompt": "What is 50% of 678, then add 49?",
"reply": "1 percent of 678 = 6.78, times 50 = 339, then 339 + 49 = 388, total 388",
"expected": 388.0,
"predicted": 388.0,
"correct": true
},
{
"source": "novel",
"task": "arithmetic",
"prompt": "Solve this basic math problem: 215 + 173",
"reply": "200 + 100 = 300, 15 + 73 = 88, total 388",
"expected": 388.0,
"predicted": 388.0,
"correct": true
},
{
"source": "novel",
"task": "percent",
"prompt": "What is 25% of 1707?",
"reply": "1 percent of 1707 = 17.07, times 20 = 341.4, times 5 = 85.35, total 426.75",
"expected": 426.75,
"predicted": 426.75,
"correct": true
},
{
"source": "novel",
"task": "average",
"prompt": "Find the average (mean) of these numbers: 22, 93, 49, 52",
"reply": "sum: 22 then 115 then 164 then 216, total 216, divide by 4, total 54.0",
"expected": 54.0,
"predicted": 54.0,
"correct": true
}
],
"non_claims": [
"Accuracy on generated arithmetic is not general problem solving. It covers five task types with small operands and nothing else.",
"A high 'seen' score with a low 'novel' score is recall, not skill; the gap is the finding, not either number alone.",
"Answer extraction takes the last number in the reply. A reply that reasons correctly and then trails off into another number scores wrong, so this is a lower bound on the model."
]
}