pino-source-code / metrics_v6_current.json
Matthew Ford
feat: harden v7 HF training and eval
3383ed6
Raw
History Blame Contribute Delete
4.76 kB
{
"checkpoint": {
"path": "models/pimt_v6.pt",
"epoch": 2,
"val_loss": 0.484091709347537,
"hidden_dim": 256,
"embedding_dim": 151,
"state_dim": 2,
"num_layers": 4,
"num_heads": 4
},
"data": {
"path": "data/empirical_dataset_v8.jsonl",
"n_records": 6078,
"n_validation_samples": 1089
},
"prediction": {
"profile_mse": 0.01670854724943638,
"profile_mae": 0.08257085829973221,
"profile_prediction_variance": 0.003912549465894699,
"tier_prediction_variance": 0.009843787178397179,
"tier_metrics": {
"top": {
"mse": 0.03479713946580887,
"mae": 0.06280406564474106,
"prediction_variance": 0.0011337018804624677,
"target_variance": 0.03625021129846573,
"mean_pairwise_cosine_similarity": 0.9688117237632773,
"mean_spearman_rho": 0.21981701217589275
},
"middle": {
"mse": 0.020692989230155945,
"mae": 0.0404493473470211,
"prediction_variance": 0.0006535974680446088,
"target_variance": 0.02135564014315605,
"mean_pairwise_cosine_similarity": 0.9567461467957046,
"mean_spearman_rho": 0.20025567720380308
},
"base": {
"mse": 0.09510858356952667,
"mae": 0.1818365752696991,
"prediction_variance": 0.021325431764125824,
"target_variance": 0.11628847569227219,
"mean_pairwise_cosine_similarity": 0.9431027251998285,
"mean_spearman_rho": 0.32991390136684196
}
}
},
"spearman": {
"flat_spearman_rho": 0.27788543250229925,
"flat_spearman_p_value": 0.0,
"profile_spearman_rho": 0.3537687699618818,
"profile_spearman_p_value": 0.0
},
"top_descriptors": {
"top": {
"predicted": [
{
"descriptor": "fresh",
"score": 0.1774865686893463
},
{
"descriptor": "sulfurous",
"score": 0.1653631627559662
},
{
"descriptor": "fishy",
"score": 0.13863913714885712
},
{
"descriptor": "fatty",
"score": 0.1347414255142212
},
{
"descriptor": "earthy",
"score": 0.130441352725029
}
],
"target": [
{
"descriptor": "fresh",
"score": 0.2029384821653366
},
{
"descriptor": "sulfurous",
"score": 0.18824610114097595
},
{
"descriptor": "fishy",
"score": 0.15518824756145477
},
{
"descriptor": "earthy",
"score": 0.1515151560306549
},
{
"descriptor": "popcorn",
"score": 0.127640038728714
}
]
},
"middle": {
"predicted": [
{
"descriptor": "fresh",
"score": 0.15246763825416565
},
{
"descriptor": "sulfurous",
"score": 0.1356164962053299
},
{
"descriptor": "grassy",
"score": 0.09715040773153305
},
{
"descriptor": "fishy",
"score": 0.07722806930541992
},
{
"descriptor": "earthy",
"score": 0.0631425753235817
}
],
"target": [
{
"descriptor": "fresh",
"score": 0.1873278170824051
},
{
"descriptor": "sulfurous",
"score": 0.15059687197208405
},
{
"descriptor": "earthy",
"score": 0.09182736277580261
},
{
"descriptor": "grassy",
"score": 0.09182736277580261
},
{
"descriptor": "fishy",
"score": 0.07346189022064209
}
]
},
"base": {
"predicted": [
{
"descriptor": "sulfurous",
"score": 0.8006615042686462
},
{
"descriptor": "fresh",
"score": 0.663232147693634
},
{
"descriptor": "fermented",
"score": 0.5910075306892395
},
{
"descriptor": "waxy",
"score": 0.5055632591247559
},
{
"descriptor": "grassy",
"score": 0.4854605197906494
}
],
"target": [
{
"descriptor": "sulfurous",
"score": 0.8126721978187561
},
{
"descriptor": "fermented",
"score": 0.6391184329986572
},
{
"descriptor": "fresh",
"score": 0.6391184329986572
},
{
"descriptor": "waxy",
"score": 0.5325987339019775
},
{
"descriptor": "balsamic",
"score": 0.5307621955871582
}
]
}
}
}