qualia_pi05 / smoke-test-v2
openpi pi05 SFT fine-tune, launched via q-research vlatrain.
- Dataset:
lerobot/aloha_sim_transfer_cube_human(LeRobot v2.1) - Steps: 200 | Batch: 32
Revisions step-<N> hold inference params only; the default revision is the final full
checkpoint (params + train_state).
Run config (verbatim, for reproducibility)
# Smoke test v2 (squashed branch) — public aloha-shaped sim dataset, minimal steps. NOT a real run.
name: qualia_pi05
exp_name: smoke-test-v2
repo_id: lerobot/aloha_sim_transfer_cube_human
checkpoint_base_dir: /ephemeral/checkpoints
project_name: qualia-pi05
hf_repo_id: qualiadev/pi05-smoke-test
data_layer: aloha
cameras: {cam_high: observation.images.top}
adapt_to_pi: false # sim data — no vendor aloha calibration
action_horizon: 16
num_train_steps: 200
save_interval: 100
keep_period: null
batch_size: 32
num_workers: 8
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: 20, peak_lr: 2.5e-5, decay_steps: 200, decay_lr: 2.5e-6}