license: other
extra_gated_prompt: TanitAD research checkpoint. Access is granted manually by the repo owner.
tags:
- tanitad
- flagship
- world-model
- autonomous-driving
- from-scratch
flagship-v4-fromscratch -- FINAL 30k checkpoint (step 29999)
The full 59.0 h (212,544.6 s wallclock) from-scratch v4 run: joint planner +
world-model training with the lambda_plan seam ramped to lam_mult_final = 1.0.
ckpt.pt is the FINAL state at final_step 29999. It is archived here because
the trainer's milestone_archives are [5000, 10000, 15000, 20000] ONLY -- there
is no 25k and no 30k milestone, so before this push the finished run existed on a
single pod disk.
MEASURED (metrics.json, in-loop, n=881 val windows)
| quantity | value |
|---|---|
final_step |
29999 |
canary_ade@2s (WM integrity) |
1.1409 |
canary_baseline |
15.6742 |
in-loop val ade@2s (DENSE-20) |
0.5063 |
in-loop val oracle_ade@2s |
0.1892 |
in-loop val miss@2m |
0.2145 |
lam_mult_final |
1.0 |
The from-scratch WM canary descended to 1.1409 against a 15.6742 baseline THROUGH full planner coupling -- that descent is the v4 thesis.
READ THIS BEFORE QUOTING A NUMBER
The in-loop ade@2s above is the trainer's DENSE-20 statistic (mean over 20
dense steps 0.1-2.0 s). It is NOT comparable to the historical
ade_0_2s/g_op_fwd_ade_m convention, which is the 4-waypoint mean over
steps 5/10/15/20. On the same forward pass at 15k this arm read dense-20 0.4596
vs 4-waypoint 0.5839. A metric NAME is not a metric DEFINITION
(TanitAD RETRACTION_LOG C1). Quote eval_flagship_v4.py output only.
ckpt_step20000.pt is the last archived milestone before the finish.
Pushed from tanitad-pod2, gated (manual approval), 2026-07-26.