Model cards

One card per reported model family. Following the Model Cards for Model Reporting convention (Mitchell et al., 2019) as it applies to a research artifact: what the model is, what it was trained on, how it was evaluated, and what it must not be used for.

Parameter counts and checkpoints are verifiable with uv run inmotion model info <name>; uv run inmotion model verify proves loadability.

World models (self-supervised)

Members: t_jepa, ts_jepa, lejepa, cf_jepa, sigreg

The mamba3_* hybrids are registered and loadable but were explored rather than reported; no metric is claimed for them.

Purpose. Learn a representation of a 10-second RSSI window without labels, then fine-tune a linear or shallow head for route classification. The premise is that interference-invariant structure is easier to learn as a prediction task than from 4-class supervision.

Architecture. A small Transformer encoder over patch/token embeddings. T-JEPA masks tabular features; TS-JEPA masks time-series patches; CF-JEPA predicts multiple horizons from short context; LeJEPA adds a Gaussian latent regulariser (SIGReg). sigreg is the regulariser as a directly supervised CNN, with no pretraining. Individual configurations are listed in docs/reference/models.md.

Training data. data/processed/dataset_augmented3.csv (35,839 rows, 13 devices), derived from dataset.csv. Augmentation: class-conditioned noise injection and intra-class mixup.

Evaluation. MCC on a held-out 20% stratified test split, seed 42. Stage-1 representation quality is tracked with a linear-probe MCC, not the pretraining loss.

Limitations.

Trained and evaluated on 13 devices from one collection campaign, in one building. Device-specific RSSI offsets are a known confound: clean_dataset.py exists because absolute signal level carried more information about which device recorded a sequence than about the route.
Leave-one-device-out performance is materially below the in-distribution number. The in-distribution MCC should not be read as cross-device generalisation.

Out of scope. Any safety-critical decision. This is a research classifier on a 4-class toy mobility task, and it is explicitly subject to device and environment shift.

Supervised deep learning baselines

Members: rnn, gru, lstm, bilstm, cnn, tcn, transformer, mamba, cnn2d_rnn, meta_fusion, plus HPO variants

Purpose. Establish what a conventional sequence model achieves on the same input, as a reference point for the world models.

Training data. data/processed/dataset.csv (3,511 rows) for the baselines; augmented variants for the HPO-best configurations.

Evaluation. Same split and metric as above.

Limitations.

The HPO-best configurations were rebuilt from documented values because the Optuna databases were deleted (inmotion train hpo-paper). The configuration is documented; the search trajectory is not recoverable.
mamba and the mamba3_* family require either the mamba_ssm CUDA extension or the pure-PyTorch fallback. The two are not numerically identical, and which one produced a given checkpoint is recorded in its log, not its weights.

Out of scope. As above.

Ensembles

Members: voting_ensemble, stacking_ensemble, deep_stack, plus the MoE variants

Purpose. Combine heterogeneous members to raise MCC. deep_stack is a 10-base stack with learned level-2 and meta models; voting_ensemble averages member probabilities with selection weights.

Combiner choices. The rationale and the literature behind them are in the module docstring of src/inmotion/pipelines/mega_ensemble.py. Briefly: Caruana-style forward selection with replacement over the member library, temperature calibration per member on the validation fold, and a regularised logistic-regression stacker as the alternative.

Evaluation. Members are combined on out-of-fold predictions to avoid optimistic selection; the ensemble is then scored once on the held-out test split.

Limitations.

The deep_stack reported artifact is backup/DeepStackEnsemble_seed42.pt (43,575,549 parameters).
Ensemble gains are computed over members trained on the same 13-device pool and do not measure robustness to a new device.

Out of scope. As above.

Classical machine learning

Members: LogisticRegression, RandomForest, ExtraTrees, GradientBoosting, XGBoost, LightGBM, CatBoost, SVC, KNN, MLP, GaussianProcess, TabPFN

Purpose. Non-neural baselines, and the models served by the demo.

Training data. Same processed datasets. TabPFN is used in inference mode over the OOF feature matrix.

Evaluation. Same split, same metric, plus cross-validation for stability.

Limitations. Standardised-RSSI models inherit the same device confound as the DL models. GaussianProcess checkpoints are ~37 MB each and are the largest classical artifacts.

Out of scope. As above.

Intended use, for every card above


Intended	research on WiFi-RSSI-based mobility classification; reproducing the reported numbers; teaching and extension
Not intended	surveillance, identification of individuals, or any operational transport decision
Privacy note	`mac` is retained to support leave-one-device-out analysis and is not a model feature, but MAC addresses are device identifiers. Treat the raw captures in `data/raw/` as personal data and check the dataset licence before redistribution.