| { |
| "architecture": "S1Model", |
| "config": { |
| "width": 128, |
| "heads": 4, |
| "enc_layers": 3, |
| "dec_layers": 2, |
| "ff": 384, |
| "dropout": 0.1, |
| "id_dropout": 0.1, |
| "pos_mode": "rand", |
| "pos_table": 48, |
| "ordinal": true, |
| "ord_table": 12, |
| "progress_head": false, |
| "invariant_numerics": true, |
| "modular": true, |
| "pointer": "done", |
| "index_eval": "identity", |
| "type_dropout": 0.15, |
| "temperature": 2.661, |
| "ui": true, |
| "ui_recent_fields": true, |
| "ui_pending_feature": true, |
| "ui_param_match": true |
| }, |
| "metadata": { |
| "name": "mesa-s1", |
| "version": "0.1.0", |
| "trained_on": "synthetic FreeCAD PartDesign episodes driven through the FreeCAD GUI (Qt widget tree)", |
| "base": "taiga-s1 architecture; SFT + DAgger on UI-level data (12k episodes, ~391k decisions + 3 DAgger rounds)", |
| "calibration": { |
| "states": 16434, |
| "episodes": 264, |
| "T": 2.661, |
| "held_out_T1": { |
| "n": 8349, |
| "acc": 0.8849, |
| "nll": 0.3749, |
| "ece": 0.0922, |
| "mean_conf": 0.9666, |
| "mean_conf_when_wrong": 0.91, |
| "n_wrong": 961 |
| }, |
| "held_out_T": { |
| "n": 8349, |
| "acc": 0.8849, |
| "nll": 0.2498, |
| "ece": 0.0457, |
| "mean_conf": 0.8845, |
| "mean_conf_when_wrong": 0.7312, |
| "n_wrong": 961 |
| } |
| } |
| } |
| } |
