doremifasolatido/pi0-fast-libero-lora

pi0-FAST LIBERO (LoRA fine-tune)

LoRA fine-tune of pi0_fast_base on the physical-intelligence/libero dataset, produced with openpi.

Training

Config (openpi)	pi0_fast_libero_low_mem_finetune
Variant	LoRA on PaliGemma (paligemma_variant="gemma_2b_lora"), EMA off
Base weights	gs://openpi-assets/checkpoints/pi0_fast_base
Dataset	physical-intelligence/libero
Steps	30,000 (save-interval=1000, max_to_keep=1 on disk)
Batch size	32
Hardware	single H200, ~13.5 h wall-clock
Final loss (step 29900)	loss=0.7988, grad_norm=6.59, param_norm=1667.07

Loss trajectory: 1.69 → 1.20 (10k) → 0.93 (20k) → 0.80 (30k).

Contents

This is the full orbax checkpoint dir at step 29999, including optimizer state — usable for both inference and resumed training.

params/             # model parameters (~5.5 GB read on restore)
train_state/        # optimizer + scheduler state (drop for inference-only)
assets/             # baked norm stats for physical-intelligence/libero
_CHECKPOINT_METADATA

Usage (openpi)

# Inside an openpi checkout:
hf download doremifasolatido/pi0-fast-libero-lora --local-dir ./pi0_fast_libero_lora_29999

XLA_PYTHON_CLIENT_MEM_FRACTION=0.4 uv run --no-sync scripts/serve_policy.py \
    --env LIBERO --port 8000 \
    policy:checkpoint \
    --policy.config=pi0_fast_libero_low_mem_finetune \
    --policy.dir=./pi0_fast_libero_lora_29999

Then drive it with examples/libero/main.py (or use this repo's scripts/run_libero_eval.py for an end-to-end harness).

Eval

LIBERO suite success rates: pending (eval running at upload time).