Behavior Uncloning — Entangled-Object KL Step 20
Experiment: Entangled-object unlearning (Pivot A) Task: Forget "milk" (T5) while retaining visually near-identical "orange juice" (T0) on LIBERO-Object Method: KL (KL Minimization) Base model: pi0.5 (3.35B params, flow matching)
Results
| Metric | Value |
|---|---|
| HM | 0.90 |
| Forget SR (milk) | 0% |
| Retain SR | 82.2% |
| OJ SR (visual neighbor) | 100% |
Context
This checkpoint is from the entangled-object forgetting experiment, which tests whether unlearning can distinguish between visually near-identical objects (milk and orange juice are both rectangular cartons of similar size).
Unlike the butter experiment where the forget target was geometrically unique, this experiment deliberately uses the hardest visual discrimination case in the dataset.
See the full report for details.
Training Config
- Suite: LIBERO-Object, Forget: T5 (milk, dataset index 5)
- Steps: 20, BS=32, LR=1e-5
- Hardware: 8xA100-80GB (FSDP)
- Framework: openpi