exp-assets (private training cache)

Pull on a fresh box instead of re-deriving. Needs an HF read token.

export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli download yaojin17/exp-assets --repo-type model --local-dir exp-assets --token <HF_TOKEN>

# HISTORY latents (f_lat=3) -> for the history runs:
mkdir -p data/navsim/trainval_hist_lat
for t in exp-assets/trainval_hist_lat_tars/*.tar; do tar xf "$t" -C data/navsim/trainval_hist_lat; done

# BASELINE latents (f_lat=2) -> for navsim_repro_cfg (original drivewam reproduction):
mkdir -p data/navsim/trainval_lat
for t in exp-assets/trainval_lat_tars/*.tar; do tar xf "$t" -C data/navsim/trainval_lat; done

# checkpoints: exp-assets/lingbot-va-base (wan22 base) , exp-assets/drivewam_navsim/transformer (init)

Contents: trainval_hist_lat (102k, f_lat=3) + trainval_lat (103k, f_lat=2) latents + lingbot-va-base + drivewam_navsim.

Eval data (navtest latents + PDM metric cache)

For the in-training PDMS eval (so the navtest dataset isn't empty):

export HF_HUB_ENABLE_HF_TRANSFER=1
# navtest latents (12,146 sample_*.pkl) -> data/navsim/test/
mkdir -p data/navsim/test
for t in exp-assets/navsim_test_tars/*.tar; do tar xf "$t" -C data/navsim/test; done
# PDM metric cache -> data/navsim/metric_cache_extract/metric_cache/
tar xf exp-assets/navsim_metric_cache/metric_cache_extract.tar -C data/navsim

Eval then uses: --dataset-path data/navsim/test --metric-cache-path data/navsim/metric_cache_extract/metric_cache