aq-mario / gates.json
maxmill's picture
Upload AQ-Mario EMA JEPA checkpoint and evaluation artifacts
e6248ba verified
Raw History Blame Contribute Delete
2.53 kB
{
"n": 24313,
"n_train_latents": 82628,
"n_val_trajectories": 60,
"x_probe_r2": 0.6882767677307129,
"y_probe_r2": 0.5903681516647339,
"scroll_probe_r2": 0.6938381791114807,
"dead_in_5_auc": 0.8556168211261861,
"dead_in_5_ap": 0.08199775871635198,
"dead_in_5_base_rate": 0.0096244802698493,
"aliasing_margin": -2.1398077743072976,
"aliasing_far_p5": 20.552905750274657,
"aliasing_near_p95": 34.21174240112305,
"aliasing_n_far": 52332,
"aliasing_n_near": 32961,
"action_effect_ratio": 0.16307152402248534,
"action_x_gap_px": 6.065155029296875,
"action_n_states": 12,
"aliasing_curve": [
{
"dx_lo": 0,
"dx_hi": 20,
"n": 32776,
"mean_dist": 23.520071029663086,
"p5": 4.397152662277222,
"p95": 34.217501640319824
},
{
"dx_lo": 20,
"dx_hi": 50,
"n": 9065,
"mean_dist": 20.05980682373047,
"p5": 10.749410247802736,
"p95": 29.08197402954101
},
{
"dx_lo": 50,
"dx_hi": 100,
"n": 13288,
"mean_dist": 23.878231048583984,
"p5": 17.462076282501222,
"p95": 29.349604892730714
},
{
"dx_lo": 100,
"dx_hi": 200,
"n": 28881,
"mean_dist": 26.2438907623291,
"p5": 20.820297241210938,
"p95": 32.452579498291016
},
{
"dx_lo": 200,
"dx_hi": 350,
"n": 26839,
"mean_dist": 26.487529754638672,
"p5": 21.805502891540527,
"p95": 31.795048713684082
},
{
"dx_lo": 350,
"dx_hi": 500,
"n": 20317,
"mean_dist": 26.41353988647461,
"p5": 21.72634963989258,
"p95": 31.10965843200684
},
{
"dx_lo": 500,
"dx_hi": 800,
"n": 25058,
"mean_dist": 26.284767150878906,
"p5": 20.94424352645874,
"p95": 31.00001220703125
},
{
"dx_lo": 800,
"dx_hi": 1500,
"n": 17448,
"mean_dist": 24.848669052124023,
"p5": 20.160357093811037,
"p95": 29.69902172088623
},
{
"dx_lo": 1500,
"dx_hi": null,
"n": 9826,
"mean_dist": 26.344118118286133,
"p5": 20.916846752166748,
"p95": 30.732340812683105
}
],
"aliasing_monotone": false,
"aliasing_spearman": 0.06707154459631581,
"aliasing_far_over_near": 1.1200696666715593,
"aliasing_usable_range": 0,
"passes": {
"x": false,
"y": false,
"scroll": false,
"aliasing": false,
"action": false
},
"aliasing_margin_strict_fails": true,
"all_pass": false,
"lemario_y_probe_r2": 0.188
}