Troubadour / SECONDARY_RESULTS.json
noahmclaughlin's picture
Release nine audited splits and completed Troubadour-Bench results
ed8341f verified
Raw History Blame Contribute Delete
13.4 kB
{
"curves": {
"base": [
{
"k": 0,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.0,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 1.0
},
{
"k": 1,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.025,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 0.975
},
{
"k": 2,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.55,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 0.45
},
{
"k": 4,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.05,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.631578947368421,
"answerable_accuracy": 0.0,
"joint_accuracy_including_uncertainty": 0.35
},
{
"k": 8,
"cohort_size": 40,
"oracle_resolvable_fraction": 1.0,
"justified_correct_resolution": 0.3,
"unjustified_certainty_rate": null,
"answerable_accuracy": 0.3,
"joint_accuracy_including_uncertainty": 0.3
}
],
"candidate": [
{
"k": 0,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.0,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 1.0
},
{
"k": 1,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.05,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 0.95
},
{
"k": 2,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.0,
"justified_correct_resolution": 0.0,
"unjustified_certainty_rate": 0.55,
"answerable_accuracy": null,
"joint_accuracy_including_uncertainty": 0.45
},
{
"k": 4,
"cohort_size": 40,
"oracle_resolvable_fraction": 0.05,
"justified_correct_resolution": 0.05,
"unjustified_certainty_rate": 0.8947368421052632,
"answerable_accuracy": 1.0,
"joint_accuracy_including_uncertainty": 0.15
},
{
"k": 8,
"cohort_size": 40,
"oracle_resolvable_fraction": 1.0,
"justified_correct_resolution": 0.625,
"unjustified_certainty_rate": null,
"answerable_accuracy": 0.625,
"joint_accuracy_including_uncertainty": 0.625
}
]
},
"rendering_checks": {
"prose": {
"base": {
"n": 48,
"accuracy": 0.3333333333333333,
"balanced_accuracy": 0.3333333333333333,
"answerable_accuracy": 0.40625,
"conditional_AB_accuracy": 0.65625,
"uncertainty_ranking_auc": 0.455078125,
"uncertainty_recall": 0.1875,
"uncertainty_precision": 0.1875,
"paired_both_correct": 0.20833333333333334,
"paired_answerable_both_correct": 0.25,
"contrast_both_queries_correct": null,
"invalid": 0,
"expected_counts": {
"U": 16,
"A": 16,
"B": 16
},
"prediction_counts": {
"A": 12,
"B": 20,
"U": 16
},
"seconds": 208.010329246521
},
"candidate": {
"n": 48,
"accuracy": 0.4166666666666667,
"balanced_accuracy": 0.4166666666666667,
"answerable_accuracy": 0.53125,
"conditional_AB_accuracy": 0.65625,
"uncertainty_ranking_auc": 0.453125,
"uncertainty_recall": 0.1875,
"uncertainty_precision": 0.23076923076923078,
"paired_both_correct": 0.2916666666666667,
"paired_answerable_both_correct": 0.375,
"contrast_both_queries_correct": null,
"invalid": 0,
"expected_counts": {
"U": 16,
"A": 16,
"B": 16
},
"prediction_counts": {
"A": 15,
"B": 20,
"U": 13
},
"seconds": 66.53535676002502
},
"society_pairs": 8,
"cluster_mean_gain": 0.08333333333333333,
"cluster_bootstrap_95ci": [
0.020833333333333332,
0.14583333333333331
],
"method": "paired bootstrap over complement-society clusters, equal cluster weighting; 10000 replicates"
},
"symbols": {
"base": {
"n": 48,
"accuracy": 0.3333333333333333,
"balanced_accuracy": 0.3333333333333333,
"answerable_accuracy": 0.5,
"conditional_AB_accuracy": 0.53125,
"uncertainty_ranking_auc": 0.6064453125,
"uncertainty_recall": 0.0,
"uncertainty_precision": 0.0,
"paired_both_correct": 0.16666666666666666,
"paired_answerable_both_correct": 0.25,
"contrast_both_queries_correct": null,
"invalid": 0,
"expected_counts": {
"U": 16,
"A": 16,
"B": 16
},
"prediction_counts": {
"A": 11,
"B": 35,
"U": 2
},
"seconds": 142.15941667556763
},
"candidate": {
"n": 48,
"accuracy": 0.4375,
"balanced_accuracy": 0.4375,
"answerable_accuracy": 0.46875,
"conditional_AB_accuracy": 0.5625,
"uncertainty_ranking_auc": 0.513671875,
"uncertainty_recall": 0.375,
"uncertainty_precision": 0.4,
"paired_both_correct": 0.16666666666666666,
"paired_answerable_both_correct": 0.1875,
"contrast_both_queries_correct": null,
"invalid": 0,
"expected_counts": {
"U": 16,
"A": 16,
"B": 16
},
"prediction_counts": {
"B": 28,
"U": 15,
"A": 5
},
"seconds": 42.30292248725891
},
"society_pairs": 8,
"cluster_mean_gain": 0.10416666666666666,
"cluster_bootstrap_95ci": [
0.0,
0.20833333333333331
],
"method": "paired bootstrap over complement-society clusters, equal cluster weighting; 10000 replicates"
}
},
"smoke": {
"base": {
"note": "12 simple instruction smoke checks, not evidence of broad general capability",
"passed": 11,
"total": 12,
"cases": [
{
"name": "addition",
"prompt": "What is 19 + 23? Reply with only the integer.",
"expected": "42",
"response": "42",
"correct": true,
"seconds": 3.0267980098724365
},
{
"name": "multiply",
"prompt": "What is 7 times 8? Reply with only the integer.",
"expected": "56",
"response": "56",
"correct": true,
"seconds": 1.733670949935913
},
{
"name": "sort",
"prompt": "Sort 9, 2, 5 in ascending order. Reply exactly as comma-separated integers without spaces.",
"expected": "2,5,9",
"response": "2,5,9",
"correct": true,
"seconds": 2.86738920211792
},
{
"name": "reverse",
"prompt": "Reverse the letters in the word lamp. Reply only with the reversed string.",
"expected": "pmal",
"response": "malp",
"correct": false,
"seconds": 2.7101151943206787
},
{
"name": "copy",
"prompt": "Reply with exactly this word: lantern",
"expected": "lantern",
"response": "lantern",
"correct": true,
"seconds": 1.4463541507720947
},
{
"name": "extract",
"prompt": "The traveler is named Mira and her companion is Oren. Reply only with the traveler's name.",
"expected": "Mira",
"response": "Mira",
"correct": true,
"seconds": 1.987515926361084
},
{
"name": "count",
"prompt": "How many items are in this list: apple, pear, plum, peach? Reply only with the integer.",
"expected": "4",
"response": "4",
"correct": true,
"seconds": 1.7788619995117188
},
{
"name": "case",
"prompt": "Convert the word TRAVEL to lowercase. Reply with only the converted word.",
"expected": "travel",
"response": "travel",
"correct": true,
"seconds": 1.897078037261963
},
{
"name": "negation",
"prompt": "The box contains a red token and no blue tokens. Does it contain a blue token? Reply exactly yes or no.",
"expected": "no",
"response": "no",
"correct": true,
"seconds": 2.0279381275177
},
{
"name": "conditional",
"prompt": "If the lamp is on, reply sun. Otherwise reply moon. The lamp is off.",
"expected": "moon",
"response": "moon",
"correct": true,
"seconds": 1.5714070796966553
},
{
"name": "format",
"prompt": "Reply with the JSON object whose only key is x and whose value is 3. Use no spaces or markdown.",
"expected": "{\"x\":3}",
"response": "{\"x\":3}",
"correct": true,
"seconds": 2.544711112976074
},
{
"name": "sequence",
"prompt": "What comes next in 2, 4, 6, 8? Reply only with the integer.",
"expected": "10",
"response": "10",
"correct": true,
"seconds": 0.9220678806304932
}
]
},
"candidate": {
"note": "12 simple instruction smoke checks, not evidence of broad general capability",
"passed": 11,
"total": 12,
"cases": [
{
"name": "addition",
"prompt": "What is 19 + 23? Reply with only the integer.",
"expected": "42",
"response": "42",
"correct": true,
"seconds": 0.4835472106933594
},
{
"name": "multiply",
"prompt": "What is 7 times 8? Reply with only the integer.",
"expected": "56",
"response": "56",
"correct": true,
"seconds": 0.49234795570373535
},
{
"name": "sort",
"prompt": "Sort 9, 2, 5 in ascending order. Reply exactly as comma-separated integers without spaces.",
"expected": "2,5,9",
"response": "2,5,9",
"correct": true,
"seconds": 0.5188479423522949
},
{
"name": "reverse",
"prompt": "Reverse the letters in the word lamp. Reply only with the reversed string.",
"expected": "pmal",
"response": "malp",
"correct": false,
"seconds": 0.632922887802124
},
{
"name": "copy",
"prompt": "Reply with exactly this word: lantern",
"expected": "lantern",
"response": "lantern",
"correct": true,
"seconds": 0.5518147945404053
},
{
"name": "extract",
"prompt": "The traveler is named Mira and her companion is Oren. Reply only with the traveler's name.",
"expected": "Mira",
"response": "Mira",
"correct": true,
"seconds": 0.7790079116821289
},
{
"name": "count",
"prompt": "How many items are in this list: apple, pear, plum, peach? Reply only with the integer.",
"expected": "4",
"response": "4",
"correct": true,
"seconds": 0.7247340679168701
},
{
"name": "case",
"prompt": "Convert the word TRAVEL to lowercase. Reply with only the converted word.",
"expected": "travel",
"response": "travel",
"correct": true,
"seconds": 0.6896908283233643
},
{
"name": "negation",
"prompt": "The box contains a red token and no blue tokens. Does it contain a blue token? Reply exactly yes or no.",
"expected": "no",
"response": "no",
"correct": true,
"seconds": 0.7926909923553467
},
{
"name": "conditional",
"prompt": "If the lamp is on, reply sun. Otherwise reply moon. The lamp is off.",
"expected": "moon",
"response": "moon",
"correct": true,
"seconds": 0.7604110240936279
},
{
"name": "format",
"prompt": "Reply with the JSON object whose only key is x and whose value is 3. Use no spaces or markdown.",
"expected": "{\"x\":3}",
"response": "{\"x\":3}",
"correct": true,
"seconds": 0.9315381050109863
},
{
"name": "sequence",
"prompt": "What comes next in 2, 4, 6, 8? Reply only with the integer.",
"expected": "10",
"response": "10",
"correct": true,
"seconds": 0.8023531436920166
}
]
}
}
}