This repo contains a minimal implementation of multi-task post-training for π0.5 on LIBERO-90 and MolmoSpaces, using filtered SFT (FSFT), Best-of-N (BoN). A OGPO implementation is on its way.
git clone git@github.com:lasgroup/generalist-improvement.git generalist-improvementEverything runs inside the container described by docker/generalist-improvement.Dockerfile.
cd generalist-improvement
docker build . -f docker/generalist-improvement.Dockerfile -t generalist-improvementTo start an interactive shell in the container, run this from the repo root.
docker run --rm -it --gpus all --network=host -v "$(pwd)":/app generalist-improvement bashThe code requires setting these variables; also consider updating global variables in scripts/launcher.py.
OPENPI_DATA_HOME = "/path/to/openpi_data_home"
MLSPACES_ASSETS_DIR = "/path/to/molmospaces/assets"
MLSPACES_CACHE_DIR = "/path/to/molmospaces/cache"
HF_HOME = "/path/to/hf_home"
HF_TOKEN = "hf_..."
WANDB_API_KEY = "..."
MUJOCO_GL = "egl"
XLA_PYTHON_CLIENT_MEM_FRACTION = "0.9"
We recommend installing assets through uv run scripts/install_molmo_assets.py, which takes a few minutes. The remaining assets are downloaded automatically.
Before launching, edit hard-coded values for your cluster and allocation at the top of scripts/launcher.py. Then, simply:
uv run scripts/launcher.py --config scripts/configs/multitask/bofn/molmo/bofn_molmo_tasks64_v0.yaml --mode local Each job requires one 4-GPU node with ~100GB VRAM each.
Note: the repo contains forks of openpi and molmospaces, which apply minimal changes on top of commits 981483dca0fd9acba698fea00aa6e52d56a66c58 and 7a2612cccdc3a7733d777b9cea1ac387d0a3db82, respectively.