TSC-FM compares foundation models for time series classification across 198 unique datasets: 128 univariate and 70 multivariate. Researchers can use the benchmark to compare pretrained models under Standard and few-shot (Low-shot) settings, inspect the effect of classifier configurations, and read the evaluation protocol behind the rankings.

This Hugging Face organization provides a research overview and a live leaderboard preview. Complete results, model documentation, capability analysis, and model registration are available on the official TSC-FM website.

Start here

• Visit the official time series classification benchmark
• Open the live Hugging Face overview
• Explore the time series classification leaderboard
• Compare foundation models for time series classification
• Read the evaluation protocol
• Submit a model for evaluation

Compare time series foundation models on classification tasks

The benchmark evaluates how pretrained models assign class labels to time series. It focuses on classification, rather than ranking forecasting performance. The unit of comparison is a complete model configuration, including its prediction approach and classifier head where applicable.

Model documentation and research resources

The catalog includes ChorusTIC, Mantis, MantisV2, MOMENT, NuTime, TabICL, TabICLv2, TIC-FM, TimEE, TiCT, and UniShape. Each method page links to available papers, code, and checkpoints and describes its published evaluation coverage. Catalog membership does not imply complete results for every setting.

In-context configurations use labeled support examples at prediction time without target-specific parameter updates. Frozen-probe configurations keep the pretrained backbone fixed and fit a classifier on its representations. TabICL and TabICLv2 are tabular foundation models applied to vectorized time series.

Evaluation scope and interpretation

Standard uses the full official training split. Few-shot results use five or ten labels per class on the eligible dataset pool for each budget. Those pools differ, so their aggregate means do not isolate the effect of adding labels. Standard and Low-shot are label-budget settings; in-context learning and frozen probes describe prediction approaches, not separate label budgets.

Compare configurations within the same setting and dataset scope, and check accuracy together with average rank and coverage. Dataset-level accuracies receive equal weight. Rankings describe the published task pools and do not establish a universally best model or statistical significance from an aggregate score difference alone. Current scores and versions remain on the official site rather than being copied into this organization card.

Paper and citation

Surveying and Benchmarking Foundation Models for Time Series Classification introduces the benchmark and the accompanying survey. It is a preprint, not a peer-reviewed publication. The paper's analyzed snapshot and the website's current results should be distinguished when citing numerical findings.

中文介绍：时间序列分类基础模型评测

TSC-FM 是面向时间序列分类基础模型的评测基准与排行榜，覆盖单变量和多变量任务，提供标准训练集、少样本分类（5-shot / 10-shot）、上下文学习与冻结特征分类器等设置的结果和解释。研究者可通过排行榜比较成绩，通过评测协议了解比较范围，并在模型目录查看论文、代码和权重来源。不同少样本预算的任务池并不完全相同，使用结果时应同时核对设置与覆盖范围。

Release and maintenance

TSC-FM is jointly released by DMIR Lab and Huawei Noah’s Ark Lab.

Maintained by Juntao Fang. For benchmark questions, contact fangjuntao1@mails.gdut.edu.cn.