Shiki42/ctr-pi05-putcab-athenb-lora-2500

CTR PyTorch PI0.5 grouped-PEFT baseline for A->B sequential.

Provenance

Experiment: E003
CTR commit: cbfb4fd15e76f729f767b9427c51487043a4d33f
RoboTwin training source commit: 6f1accb89772de393946112c73e79ea9f593c866
Dataset: Shiki42/robotwin_putcab_fixed_leftDrawer_rightObject_AthenB_50
Dataset revision: 793556b0d670af13653dcb6a4ba99864fd985c59
Base model SHA-256: d34fa1553e4321c5578cded8fc0b95e98859b4646525fc9ed52a24cbb5365741
Runtime manifest SHA-256: 06e035fbc62dd7f32f01a2f5ad536f01bd5af6caae3239698dbb8c07ed503332
PEFT version: 0.17.1

Optimizer updates: 2500
Micro batch: 8
Gradient accumulation: 2
Effective global batch: 16
Sample budget: 40000 rows
Seed: 87431
Precision: bfloat16
EMA: None
LoRA recipe: {"expert_alpha": 32.0, "expert_rank": 32, "language_alpha": 16.0, "language_rank": 16}

Status: PASS
Success: 5/20 (25.0%)
Collision: 4/20 (20.0%)
Ordered seed manifest SHA-256: eb93c8dd7e2ed5ffc73c7e7d95e07b2b8ab57ee2a7342c166a063c5799acec57
Rollout metrics SHA-256: 61c7258309eef5b9817ee5f0274d8b6f69ac4d7f8bec7135a5bdef502fcdcea7

This checkpoint is an experimental RoboTwin policy artifact. It is intended for research evaluation.

Video Preview