BridgeQuant-Async: LIBERO eval pack (2026-11-10)
Checkpoints awaiting LIBERO rollouts, laid out so that downloading into runs/ of the
bridgequant-async repo (commit 5236626 or later) reproduces runs/<run>/checkpoint_last.
hf download twanghcmut/bridgequant-libero-eval-pack --local-dir runs
bash runs/_pack/run_eval_pack.sh # resumable; groups: probes | full | h50 | frozen | ternary | ternary_scratch
| group | run dir | config (configs/bridgequant_vla/experiments/) |
evals |
|---|---|---|---|
| probes | probe_spatial_* (7) |
<run>.yaml |
LIBERO-Spatial 500 eps, K=16, noise seed 0 |
| full | joint4_tinyvl_tbexcess_augS_wd3e-3_full |
tinyvl_joint4_tbexcess_augS_wd3e-3_full.yaml |
4 suites x 500 eps x 3 noise seeds, K=16 |
| h50 | joint4_tinyvl_tbexcess_augS_wd3e-3_H50 |
tinyvl_joint4_tbexcess_augS_wd3e-3_H50.yaml |
same, plus one K=25 pass |
| frozen | joint4_tinyvl_tbexcess_augS_wd3e-3_frozen |
tinyvl_joint4_tbexcess_augS_wd3e-3_frozen.yaml |
same as full |
| ternary | joint4_nano_theia_w128_kd_ternary |
nano_joint4_theia_w128_kd_ternary.yaml |
same as full |
_pack/egl_ok.py is a pixel gate: it exits non-zero when the simulator renders black frames.
Throughput is CPU-bound (LIBERO stepping and rendering); success rate does not depend on it.