Hugging Face overview · Live leaderboard preview
Time Series Classification Foundation Models | TSC-FM
Compare time series classification foundation models with TSC-FM, a research benchmark spanning 220 univariate and multivariate datasets. The Standard leaderboard reports 17 complete model configurations and 3,740 configuration–dataset results. Explore the live rankings, model documentation and separate few-shot evaluations.
What this leaderboard compares
TSC-FM evaluates foundation models for time series classification on 140 univariate and 80 multivariate datasets using their default training/test splits. Standard and Low-shot results use separate label budgets; in-context prediction and frozen-probe evaluation describe how each configuration makes predictions. The live Top 5 below offers a quick view, while the full leaderboard provides accuracy, average rank, coverage, and evaluation settings.
Connecting to the official benchmark…
Live benchmark data is temporarily unavailable.
No stored results are shown, so this Space cannot drift from the official leaderboard.
Open the official leaderboard ↗
Standard Overall · Accuracy
Live foundation model leaderboard
Equal weight per dataset
| Rank | Model | Avg. accuracy | Avg. rank | Coverage |
|---|
Evaluation settings and approaches
Standard and Low-shot define label budgets. ICL and Frozen-Probe describe prediction approaches and can overlap with those settings.
Read the protocol ↗Compare where models excel
—configurations
—backbones
Foundation models for time series classification
The complete Standard Overall comparison includes ChorusTIC, Mantis, MantisV2, MOMENT, UniShape, NuTime, TabICL and TabICLv2. TabICL and TabICLv2 are tabular foundation models evaluated on vectorized time series. Several backbones have multiple classifier heads, so a configuration and a backbone are different units of comparison.
Use the method pages to inspect papers, code, checkpoints and classifier settings. Low-shot results use separate eligible task pools; their coverage should not be inferred from the Standard dataset count.
Univariate and multivariate time series classification
Time series classification assigns a class label to an observed sequence. Univariate tasks contain one measurement channel; multivariate tasks contain several channels. TSC-FM reports these groups separately as well as together, so researchers can check whether an overall result also holds for their task type.
The dataset profile explains the composition of the benchmark. The capability map examines performance across task attributes using its documented comparison core. These views help turn an overall ranking into a more specific model-selection question.
Explore classification datasets ↗ Compare model capabilities ↗Few-shot time series classification: 5-shot and 10-shot
Standard evaluation uses the full default training split. Low-shot evaluation uses five or ten labeled examples per class on the eligible dataset pool for each budget. These pools differ, so a change in their aggregate accuracy does not isolate the benefit of additional labels.
In-context learning (ICL) uses labeled support examples at prediction time. Frozen-probe evaluation keeps the pretrained backbone fixed and fits a classifier on its representations. These are prediction approaches; Standard and Low-shot describe the available label budget.
Read the few-shot evaluation protocol ↗How to compare time series foundation models
- Match the setting and scope. Select the label budget and the univariate, multivariate, or overall view relevant to your task.
- Read accuracy, average rank, and coverage together. Dataset-level accuracies receive equal weight. Check which datasets have results before drawing conclusions from an aggregate score.
- Inspect the complete model configuration. The prediction approach and classifier head can affect the comparison. The model catalog documents configurations and links to available papers, code, and checkpoints.
The live preview summarizes published results; it does not establish a universally best model. Consult the evaluation protocol and task-specific evidence before selecting a model for a new application.
Research behind the TSC-FM benchmark
Surveying and Benchmarking Foundation Models for Time Series Classification presents the accompanying survey and benchmark. It is a preprint and has not been peer reviewed.
Distinguish the paper's analyzed snapshot from the current website results when citing numerical findings. The official leaderboard provides the current results, and the evaluation page explains their scope and limitations.
时间序列分类基础模型评测与排行榜
TSC-FM 覆盖单变量和多变量时间序列分类任务，提供标准训练集与少样本分类（5-shot / 10-shot）的结果，并解释上下文学习与冻结特征分类器等预测方式。研究者可结合准确率、平均排名和数据集覆盖范围比较模型配置。
此 Hugging Face 页面提供评测介绍和实时排行榜预览。完整结果、模型论文与代码来源、评测协议及模型提交入口均可在官方站点查看。
查看完整分类排行榜 ↗