tags:
- model_hub_mixin
- pytorch_model_hub_mixin
CleanTS Model Card
Introduction
CleanTS is a competitive pre-trained baseline for time-series forecasting. It is designed to demonstrate the potential of data-centric optimization using a standard encoder-only Transformer architecture without any structural modifications.
Key Technical Principles
1. Pure Architecture
Zero modifications to the attention mechanism or model structure. CleanTS proves that high-quality data governance can significantly elevate the performance of vanilla architectures.
2. Systematic Data Governance
CleanTS adheres to a strict data-centric pipeline:
- Zero Synthetic Data: All training is performed exclusively on authentic, real-world data.
- Publicly Sourced: The training corpus consists entirely of publicly available datasets, ensuring transparency and accessibility.
- Advanced Cleaning: We achieve promising results solely through systematic data cleaning and preprocessing strategies.
3. Contamination Prevention
We strictly guarantee that no data from the GiftEval test set，Fev-bench，Fev-leaderboard and LSTF was involved in the training phase.
A detailed technical report, including our specific data cleaning methodologies, training configurations, and comprehensive ablation studies, will be released concurrently with the upcoming publication of the full model.