pi05-qua1507-restockshelfs

qualia_pi05 / qua1507_restockshelfs_10k

openpi pi05 SFT fine-tune, launched via q-research vlatrain.

• Dataset: qualiadev/qua1507-restockshelfs-success (LeRobot v2.1)
• Steps: 10000 | Batch: 32

Revisions step-<N> hold inference params + assets (norm stats) in the same layout as the default revision, which is the final full checkpoint (adds train_state).

Run config (verbatim, for reproducibility)

# pi05 SFT on qualiadev/qua1507-restockshelfs-success (OpenArm bimanual, 16-DOF).
# 95 episodes / 18,784 frames @ 20fps -> 10k steps @ batch 32 ~= 17 epochs.

name: qualia_pi05
exp_name: qua1507_restockshelfs_10k
repo_id: qualiadev/qua1507-restockshelfs-success
checkpoint_base_dir: /ephemeral/checkpoints
project_name: qualia-pi05
hf_repo_id: qualiadev/pi05-qua1507-restockshelfs

# ── Data: OpenArm bimanual (generic layer) ──────────────────────────────────────────────────────────
data_layer: generic
cameras:
  cam_1: observation.images.context
  cam_2: observation.images.wrist_left
  cam_3: observation.images.wrist_right
relativize: true
bimanual: true

# ── Model ──────────────────────────────────────────────────────────────────────────────────────────
action_horizon: 16

# ── Hyperparameters ────────────────────────────────────────────────────────────────────────────────
num_train_steps: 10000
save_interval: 2500
keep_period: null
batch_size: 32
num_workers: 16
seed: 42
fsdp_devices: 1
ema_decay: null
overwrite: false
resume: false
optimizer:
  b1: 0.9
  b2: 0.95
  eps: 1.0e-8
  weight_decay: 1.0e-10
  clip_gradient_norm: 1.0
lr_schedule:
  warmup_steps: 1000
  peak_lr: 2.5e-5
  decay_steps: 10000
  decay_lr: 2.5e-6