neurips2026citympc/citympc-txvit

CityMPC — TX_ViT pre-trained checkpoints (5 cities)

Stripped Lightning checkpoints for the CityMPC TX_ViT model from the NeurIPS 2026 paper. One checkpoint per city, selected by the seed-sweep stability filter (Kendall weighting σ_presence < 1e-3 + good-verdict on test-set outlier rule), then ranked by lowest Path Loss MAE / best Presence F1.

Per-city metrics (test set)

City	Path Loss MAE (dB)	ToF MAE (ns)	Avg Delay MAE (ns)	Presence F1	Seed
tx_austin	1.107	5.18	16.36	0.894	2024
tx_dallas	1.451	5.29	9.48	0.879	123
tx_denver	1.976	3.98	10.15	0.863	123
tx_fortworth	1.736	4.07	11.13	0.881	123
tx_newyork	2.481	8.59	19.96	0.874	42

How to use

These checkpoints are stripped — inference only, not training-resumable. After cloning the CityMPC repo:

# download a single city
hf download neurips2026citympc/citympc-txvit \
  outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt \
  outputs/tx_austin/city_10_austin_3p5_s/final/config.yaml \
  --local-dir .

# evaluate
python scripts/analyze.py \
  --ckpt outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt

Or open notebooks/02_analyze_trained_model.ipynb and paste the downloaded run path into TXVIT_RUN_DIR.

File layout

outputs/
  tx_austin/city_10_austin_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_dallas/city_8_dallas_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_denver/city_18_denver_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_fortworth/city_12_fortworth_3p5_s/final/{ckpts/last.ckpt, config.yaml}
  tx_newyork/city_0_newyork_3p5_s/final/{ckpts/last.ckpt, config.yaml}

Integrity (SHA-256)

File	Size	SHA-256
outputs/tx_austin/city_10_austin_3p5_s/final/ckpts/last.ckpt	60.6 MB	2c4b31591eb347caab9acf11623dcc2d8a184515fbf8c4186b4958a2c5fe95a3
outputs/tx_austin/city_10_austin_3p5_s/final/config.yaml	0.0 MB	f3e75a17d3d7726cfd01a75df635b26aac3ca54edd386af2b604eaaa7e3057c1
outputs/tx_dallas/city_8_dallas_3p5_s/final/ckpts/last.ckpt	60.6 MB	397327907665bb7abb70335a0cd581677a8b6b55e9d3a4fbe95092d7fd9d12ad
outputs/tx_dallas/city_8_dallas_3p5_s/final/config.yaml	0.0 MB	9b60b5483f2580737052ba8ed6e4d89fe86dd52161095b50a14afcbe7256a88d
outputs/tx_denver/city_18_denver_3p5_s/final/ckpts/last.ckpt	60.6 MB	13e496219b7210f9092996023f16ef0aed53b066ac46b322cf8b664085706dee
outputs/tx_denver/city_18_denver_3p5_s/final/config.yaml	0.0 MB	b03ea4570bbd423a719f91919293ed23617d48476476f7f398fb1ab0d3f51ea0
outputs/tx_fortworth/city_12_fortworth_3p5_s/final/ckpts/last.ckpt	60.6 MB	6f6ce2e43b025edf422de5b94edc8c476dbc4d604137d487f00cf494c88f9317
outputs/tx_fortworth/city_12_fortworth_3p5_s/final/config.yaml	0.0 MB	d7d124c9ff9d28c4ac8c7da2c3273863a4466e567480fae382593b5c9bd32e34
outputs/tx_newyork/city_0_newyork_3p5_s/final/ckpts/last.ckpt	60.6 MB	98de5771b763445dd04d8c0bdf070777492710ffb0e65fa7807e599871b00e74
outputs/tx_newyork/city_0_newyork_3p5_s/final/config.yaml	0.0 MB	c33baa28a3d6ae02626a46cd64f3161378cb53e4d93968528cc46091c861065d

Notes

• Architecture: TX_ViT conditional VAE (CityMPC). See the paper for details.
• Trained on the CityMPC dataset.
• Lightning state stripped (optimizer / scheduler / callback) — checkpoints are not training-resumable.
• σ-gate: every shipped run had weighting/sigma/presence < 1e-3 at its last logged step.