Update card to iter-5800 (2xH100 DDP resume) + refreshed eval

README.md

@@ -11,8 +11,8 @@ tags:
 # CAC — from-scratch V-JEPA 2-AC predictor
 
 A from-scratch reproduction of Meta's V-JEPA 2-AC action-conditioned predictor,
-trained on a single H100. This repo holds only the trained **predictor**; the
-frozen V-JEPA 2 ViT-g encoder is not included — use Meta's `vjepa2_ac_vit_giant`
+trained on H100s. This repo holds only the trained **predictor**; the frozen
+V-JEPA 2 ViT-g encoder is not included — use Meta's `vjepa2_ac_vit_giant`
 (torch.hub) for the encoder.
 
 ## Architecture
@@ -24,14 +24,17 @@ action + state token conditioning (7-dim each). Identical to the shipped AC.
 
 ## Training (this snapshot)
 
-- **2,750 iterations** (toward Meta's 94,500-iter recipe; stopped early to
-  benchmark), effective batch 256, LR 7.5e-5 → 4.25e-4 → 0 (WSD), AdamW wd 0.04,
-  bf16, `torch.compile`.
+- **5,800 iterations** (toward Meta's 94,500-iter recipe; a training snapshot,
+  not the finished run), effective batch 256, LR 7.5e-5 → 4.25e-4 → 0 (WSD,
+  scheduled over the full 94.5k horizon), AdamW wd 0.04, bf16, `torch.compile`.
 - Loss: L1 teacher-forcing + 2-step autoregressive rollout (faithful port of
   Meta's `app/vjepa_droid/train.py`).
 - Data: ~51,000 DROID left-cam episodes (~58 h footage — essentially the full
   DROID 1.0.1 raw corpus), 8 frames @ 4 fps, 256 px; first 240 held out for eval.
-- Hardware: 1× H100 80GB (GCP spot), ~9 h.
+- Hardware: 1× H100 80GB to iter 2,800, then resumed 2,800 → 5,800 on **2× H100
+  DDP** (5.82 s/iter, 1.98× single-GPU — near-linear over NVLink). Resume was
+  seamless: no loss spike across the boundary, the WSD LR schedule continued on
+  the 94.5k horizon, and val cosine ticked 0.818 → 0.823.
 
 ## Evaluation (240 held-out DROID episodes, frozen ViT-g)
 
@@ -39,18 +42,19 @@ Two metrics, vs Meta's shipped V-JEPA 2-AC on the same set:
 
 | metric | CAC (this) | V-JEPA 2-AC |
 |---|---|---|
-| prediction cosine | **0.782** | 0.766 |
-| CEM plan L2 → true action | 0.055 | **0.043** |
-| cos(plan, true action) | +0.58 | **+0.66** |
-| CEM beats random | 86% | **94%** |
+| prediction cosine | **0.789** | 0.766 |
+| CEM plan L2 → true action | 0.048 | **0.043** |
+| cos(plan, true action) | +0.54 | **+0.66** |
+| CEM beats random | 85% | **94%** |
 
 CAC **beats** AC on prediction cosine (the encoder-dominated metric, where the
-~51k-episode data scale helps). AC **beats** CAC on planning (the dynamics-
-quality metric that needs iter count — Meta trained ~34× longer). CAC still
-plans meaningfully: it beats a random action on 86% of episodes and its plan L2
-(0.055) beats a zero/static-action baseline (0.069), so it learned
-action-conditioned dynamics, not just "hold still." Prediction cosine is a
-lenient metric; planning quality is the real test.
+~51k-episode data scale helps). On planning, CAC's plan L2 (0.048) is now **on
+par** with AC (0.043); direction still lags (+0.54 vs +0.66) — the axis bound by
+iter count, and Meta trained ~16× longer. Open-loop L1 prediction now **beats a
+static baseline at every horizon** (the undertrained iter-2,750 snapshot could
+not), CAC points the right direction on 82% of episodes, and its CEM plan beats
+a random action on 85%. Prediction cosine is a lenient metric; planning quality
+is the real test.
 
 ## File