djkesu/trlc-mixed10-basepi05-step1500-v1

OpenPI checkpoint exported from Modal training.

Checkpoint

• config: pi05_trlc_mixed10
• experiment: trlc_mixed10_basepi05_20260324_a100x8_bs256_fast20k_val1pct_v1
• uploaded checkpoint step: 1500
• variant: custom
• includes train_state: true

Dataset

• dataset: Gongsta/trlc_tshirt_folding,Gongsta/trlc_tshirt_folding_impedance,Gongsta/dagger_dk1_tshirt_corrections
• default prompt: "Pick up the t-shirt, stretch out the corners to lay it flat, and smooth out any wrinkles. Perform precise, sequential folds to condense it into a neat, perfectly aligned rectangle, matching the quality of a professional laundromat."

Files

• params/: inference-ready OpenPI parameters
• assets/: normalization stats required by OpenPI serving
• _CHECKPOINT_METADATA: checkpoint metadata
• checkpoint_info.json: upload metadata
• train_state/: resume state for further training

W&B

• run: trlc_mixed10_basepi05_20260324_a100x8_bs256_fast20k_val1pct_v1
• url: https://wandb.ai/km67000-palatial/openpi/runs/u7vxr7qz
• run val/best_loss: 0.10333611071109772
• run final step: 1400

Usage

Download the repo locally, then serve it with the OpenPI TRLC wrapper:

cd /workspace/openpi
uv run /workspace/recap/scripts/serve_trlc_openpi.py --variant custom --checkpoint-dir /path/to/downloaded/repo --port 8000