Fix speed-control model card

README.md

@@ -1,25 +1,27 @@
 ---
 license: other
-library_name: jax
+library_name: pytorch
 tags:
   - robotics
-  - preference-optimization
-  - flow-dpo
+  - reinforcement-learning
+  - rainbow-dqn
+  - speed-control
   - metarobobench
-  - pi05
 ---
 
-# MetaRoboBench Pi05 Flow-DPO Checkpoints
+# MetaRoboBench Speed-Control Heads
 
-Private offboarding archive of complete Flow-DPO parameter trees and training metrics.
+Private offboarding archive of Rainbow DQN, CQL, and candidate-value heads trained for chunk-level Fast/Normal/Slow selection on top of Pi05 checkpoint `21650`.
 
-Available complete runs:
+The primary model is `rainbow/shared_all50_final/models/shared_all50.pt`:
 
-- `checkpoints/g16_s500_conservative`: completed step 500
-- `checkpoints/g4_s3000_b01_a005`: completed step 3,000
-- `checkpoints/g8_s1000_b01_a005_r2`: completed step 1,000
+- 50-task shared Rainbow head
+- 100,491 decisions and 99,962 updates
+- 1,941/2,500 successes = 77.64%
+- Fast/Normal/Slow chunk shares = 38.6/32.3/29.1%
+- Mean speed switches = about 4.23 per episode
 
-These runs were initialized from the frame-aligned Pi05 SFT family. They are research checkpoints, not the main fixed-speed performance baseline. Incomplete 20,000-step and pilot runs are excluded because no full parameter tree remains.
+The accelerated shared checkpoint is a distinct earlier checkpoint, not a duplicate. Per-task Rainbow, CQL, and candidate-value models are retained as baselines and negative-result artifacts.
 
 Code reference: