hilserl EXPO-FT learner environment (A100 host, 2026-10-09)
Snapshot of everything installed for the pi0.5 EXPO-FT learner (hilserl.algos.expo_ft.learner --sft-only) on
an A100 80GB x2 host (Ubuntu 22.04, driver 535.183.06), for restoring the same host layout (HOME=/home/elicer).
| file | extracts to | contents |
|---|---|---|
expo-ft.tar |
/home/elicer/expo-ft |
github.com/pd-perry/expo-ft 21fe3d3 + openpi fork (branch expo_ft) 46407a4 + .venv (uv sync, JAX 0.5.3, + tensorboardX 2.6.5, msgpack-numpy 0.4.8) |
uv.tar |
/home/elicer/.local/{bin,share/uv/python} |
uv 0.12.24 and CPython 3.11.17 the venv links to |
openpi-cache.tar |
/home/elicer/.cache/openpi |
openpi-assets/checkpoints/pi05_base/params, big_vision/paligemma_tokenizer.model |
Restore: bash restore.sh <download dir>. Not included: the Hugging Face token (run hf auth login), the
agentic-behavior-1k checkout (clone branch hil-serl; its hilserl/setup_learner_expo_ft.sh sets the host
values and also builds this environment from scratch on any path).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support