README.md

If this resource has been helpful to you, we'd love it if you could give us a star ⭐ and cite our research!

@inproceedings{
hu2025contextalignment,
title={Context-Alignment: Activating and Enhancing {LLM}s Capabilities in Time Series},
author={Yuxiao Hu and Qian Li and Dongxiao Zhang and Jinyue Yan and Yuntian Chen},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=syC2764fPc}
}

The official implementation of our ICLR-2025 paper "Context-Alignment: Activating and Enhancing LLM Capabilities in Time Series" [Arxiv] [ICLR]. The repo using Wandb to manage.

Introduction

Recently, leveraging pre-trained Large Language Models (LLMs) for time series (TS) tasks has gained increasing attention, which involves activating and enhancing LLMs’ capabilities. Many methods aim to activate LLMs’ capabilities based on token-level alignment, but overlook LLMs’ inherent strength in natural language processing — their deep understanding of linguistic logic and structure rather than superficial embedding processing. We propose Context-Alignment (CA), a new paradigm that aligns TS with a linguistic component in the language environments familiar to LLMs to enable LLMs to contextualize and comprehend TS data, thereby activating their capabilities.

Usage

All experiments were performed using NVIDIA H800 80GB GPUs, NVIDIA A800 80GB GPUs or GeForce RTX 4090 GPUs. (The experiments can be run on a single GPU, but please note that some settings may require more than 24GB of VRAM.)

1. Prerequisites:

   Note that FSCA is only tested on Ubuntu OS with the following environments. It may work on other operating systems (i.e., Windows) as well but we do not guarantee that it will.

   • Please refer to env.sh and requirements.txt to complete the environment configuration

2. Prepare the datasets:

   • Download all the datasets from [TimesNet].

3. Training Configuration:

   • For each task, you can use the subfolder scirpts for training, and the training results will be visualized and recorded via Wandb.
   • Wandb records of long-term forecasting and zero-shot forecasting are provided for you to refer.

Acknowledge

We appreciate the following github repos a lot for their valuable code base or datasets: [Time-LLM][One Fits All][TimesNet]