Mantis
Summary
Mantis is a lightweight time-series classification foundation-model lineage. In this wiki it covers the original Mantis-8M contrastive model, MantisPlus and MantisV2 synthetic-data variants, and UTICA’s self-distillation branch on the Mantis-style backbone.
Lineage
- Mantis introduces Mantis-8M, a calibrated classification-oriented encoder with a time-series token generator, ViT-style transformer, frozen-feature and fine-tuning evaluations, and multivariate adapters.
- MantisV2 extends the line with synthetic CauKer-style pretraining, a refined smaller encoder, intermediate-layer feature choices, token aggregation, self-ensembling, and feature fusion.
- UTICA keeps the Mantis tokenizer/backbone but changes the pretraining objective to DINO/iBOT-style self-distillation with masked patch prediction and KoLeo regularization.
Official Artifacts
- Mantis source and MantisV2 source: https://github.com/vfeofanov/mantis
- Mantis-8M checkpoint: https://huggingface.co/paris-noah/Mantis-8M
- MantisPlus checkpoint: https://huggingface.co/paris-noah/MantisPlus
- MantisV2 checkpoint: https://huggingface.co/paris-noah/MantisV2
- UTICA source: https://github.com/fegounna/Utica
- UTICA checkpoint: https://huggingface.co/fegounna/Utica
Role In The Wiki
Mantis is the central classification lineage for comparing contrastive pretraining, synthetic-data pretraining, test-time representation choices, and non-contrastive self-distillation in time-series classification.