vla_real_arr_bbr_gradgate_step400
Edited pi0.5 VLA checkpoint for arrangement task — arr_bbr_gradgate arm at step 400.
Deployment goal
Bias model toward the black-red-blue (BBR) arrangement order over the other two valid arrangements. Useful when the deployment scenario prefers this specific ordering.
Edit recipe
- steering_mode:
hidden_v8_mc_allpairs_precommit_gated
- target_subset:
1 (single target — black-red-blue)
- loss_formula:
-z[target] + logsumexp(z_others) (anti-saturation, gradient flows into all competing logits)
- gating: GRADIENT GATE — top-10% of unpref frames (other 2 modes' episodes) by
|∂(v8_loss)/∂h|, computed offline against the frozen mixed VLA + v5h-mc classifier
- gate_calibration: 1268 of ~13,000 unpref frames retained (~10%); concentration: EARLY-MID (more balanced, 21% in 0-30, 22% in 60-90)
- ablation_arm: GRAD-GATE (experimental — adapts per-episode to where mode commitment is happening)
Common hyperparameters
γ=0.1, β=1.0, lr=1e-5, batch=32, num-steps=400, save-interval=100, ViT frozen.
Pref episodes: 30-59 (target mode).
Unpref episodes: the other 60 (other two modes).
Foundation VLA
pi05_real_arr_mixed/real_arr_mixed_v3 step 24999 — frozen mixed-mode foundation. ViT frozen during editing; LLM + action expert updated.
Classifier used
/mnt/data3/classifiers/real_v3/arr_v5h_mc/best_v5h_mc.pt — v5h-mc 3-class softmax classifier on hidden states (val_acc 98-100%, very confident discriminator).
Why this checkpoint?
Top-2 by composite score: val_loss_pref + 0.1 × loss_redirect (lower=better).
The composite balances target-mode preservation (val_loss_pref low) with active editing pressure (loss_redirect strongly negative).
For ALL 3 arr_gradgate edits, the LATEST ckpts (step 400 and 300) won — editing benefits accumulate throughout training.
Note on gradient gate for arr-style tasks
Unlike pk_remove_sharp (where gradient peaks in EARLY frames where classifier is in transition), arr modes have classifier-confidence saturation that peaks LATE in trajectories (after items are placed in mode-specific order). The gradient gate selects post-commit frames where the editing pressure is strongest. This is a different semantic than pre-commit gating — it's "behavioral surgery" on already-committed states.
Empirical question this checkpoint helps answer: does post-commit grad-gate editing actually redirect deployment behavior, or do the LLM forward dynamics resist it? Companion ablation arms (arr_brb_full, arr_brb_early) test against full-trajectory and pre-commit-frame variants.
Ablation companion
arr_brb_full — same target, gating = frame_index < 999 (entire trajectory, no gating). Control.arr_brb_early — same target, gating = frame_index < 24 (first 10% of trajectory, sim 3pillar locked semantic).arr_bbr_gradgate (this checkpoint) — gating = top-10% by gradient magnitude, post-commit.
Eval all to validate the gating choice.
Eval target
50-seed real-robot rollouts. Compare:
- Target arrangement rate: fraction of episodes where items end up in black-red-blue order
- Overall task SR: model still completes the arrangement task
- vs foundation VLA baseline (no editing)
