Downloads · 30 days
0
pusht-v2-jcont
Sumo00048/pusht-v2-jcont
pusht-v2-jcont is a machine learning model from Sumo00048. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Continuous-reasoning-latent (Jcont) checkpoints for the Push-T v2 J-vs-C study. Load with policyfromconfigcont (repo github.com/BabaYaga840/steeringjointobs, branch worktree-pusht-v2-humanlike); the two baselines/.pt…
Access
Public
Updated Sep 6, 2026
Repo size
2.8 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt3.3 GB · 100%
At a glance
- Access
- Public
- Created
- Sep 6, 2026
- Updated
- Sep 6, 2026
- SHA
- 7ada866a
Model card
From the Hugging Face model README
Push-T v2 -- Jcont "loose conditioning" weights
Continuous-reasoning-latent (Jcont) checkpoints for the Push-T v2 J-vs-C study.
Load with policy_from_config_cont (repo github.com/BabaYaga840/steering_joint_obs,
branch worktree-pusht-v2-humanlike); the two baselines/*.pt are discrete PTJointPolicyV2/
PTCondPolicy -- load those with pt_eval_v2.load_policy_v2. Full guide, inference contract,
and results: docs/pusht_v2/HANDOFF.md in that branch.
jcont_fullcov_step1/-- step-1, full coverage (mechanism proof; tight hold 0.725)jcont_sev2_hole/-- step-2, severity-2 hole (recovery NEGATIVE, graceful-degradation POSITIVE)jconts_strengthened/-- cond_every_step variant (did not lift the ~0.70 ceiling)baselines/-- discrete J & C at the sev-2 hole, seed 0 (references)
Each Jcont dir: selected.pt (inference), best_val_flow_clean.pt/final_ema.pt (candidates),
latest.pt (full training state incl. optimizer -- for resuming). All results are seed 0.
- Created
- Sep 6, 2026
- Updated
- Sep 6, 2026