# Mantis reproduction bundle

Reproduction of **Mantis: Lightweight Foundation Model for Time Series Classification**
(arXiv:2502.15637, OpenReview gbJMAjXLZ4) for the HF "Reproducing ICML 2026" challenge.

Logbook: https://huggingface.co/spaces/ancs21/repro-mantis
Paper code: https://github.com/vfeofanov/mantis  ·  Weights: https://huggingface.co/paris-noah/Mantis-8M

- `param_count.py` — Claim 1: Mantis-8M has 8,103,936 params (8.10M), matching "8M lightweight". Verified.
- `zeroshot_probe.py` — Claim 2: frozen embeddings + logistic regression on 4 UCR datasets,
  mean accuracy 0.938 (well above chance). Frozen zero-shot features are useful. Verified.

Install: `pip install mantis-tsfm aeon scikit-learn torch`. Inference-only, CPU, minutes.