NetManAIOps/ChatTS

What it solves

ChatTS is designed to enable Large Language Models (LLMs) to natively understand and reason over time series data. Unlike traditional approaches that might treat time series as text or images, ChatTS allows users to interactively explore the structure, changes, and numerical values of multivariate time series through a conversational interface.

How it works

ChatTS is a Time Series Multimodal LLM (TS-MLLM). It processes multivariate time series and text queries separately before combining them. Time series data is split into fixed-size patches and encoded via a dedicated time series encoder. These encoded patches are then concatenated with text tokens at the token level, preserving the original input positions and contextual relationships. This allows the LLM to perform unified multimodal understanding and response generation.

Who it’s for

Data scientists, AI researchers, and analysts who need to perform complex reasoning and question-answering over time series data without relying on separate prediction or classification tools.

Highlights

• Native Multivariate Support: Handles multiple time series of varying lengths and dimensions in a single prompt.
• Value-Preserved Encoding: Retains raw numerical values, allowing the model to answer specific questions about spikes or values at particular timestamps.
• High Efficiency: Outperforms GPT-4o on numerical alignment tasks while using fewer tokens.
• Flexible Deployment: Supports integration with Transformers and vLLM, including an OpenAI API-compatible server option.
• Synthetic Data Pipeline: Includes a full pipeline for generating alignment and SFT (Supervised Fine-Tuning) datasets to train similar models.