ChatTS is a multimodal large language model designed for understanding, chatting, and reasoning about time series data. It targets data scientists and researchers who need to interactively explore and gain insights from time series, offering a conversational interface for complex analysis.

How It Works

ChatTS is built natively for time series as a core modality, enabling flexible input of multivariate time series with varying lengths and dimensions. It preserves raw numerical values, allowing for precise statistical queries. The model leverages a synthetic data generation pipeline (TSEvol) and is fine-tuned on a modified QWen2.5-14B-Instruct base model, facilitating conversational understanding and reasoning over time series data.

Quick Start & Requirements

• Install: pip install -r requirements.txt (includes deepspeed, vllm==0.8.5, torch==2.6.0, flash-attn).
• Prerequisites: GPU with sufficient memory (A100/A800 recommended), CUDA, Python >= 3.11. Flash-Attention is essential.
• Setup: Download model weights from HuggingFace and place under ckpt/. Download evaluation datasets from Zenodo and place under evaluation/dataset/.
• Resources:
  • Paper: arXiv:2412.03104
  • Model: Hugging Face
  • Datasets: Hugging Face, Zenodo
  • Demos: demo_hf.ipynb, demo_vllm.py

Highlighted Details

• Native support for multivariate time series with flexible input lengths and dimensionality.
• Enables conversational interaction for time series exploration and reasoning.
• Preserves raw numerical values for accurate statistical analysis.
• Supports vLLM for efficient inference, with experimental integration available.
• Offers tools for generating synthetic time series data and training datasets.

Maintenance & Community

• The project is associated with Bytedance Research.
• Updates include new quantized models (GPTQ-4bit), data generation code, and baseline model implementations.
• Training scripts are available separately at ChatTS-Training.

Licensing & Compatibility

• Licensed under the MIT License.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The model is recommended for time series lengths between 64 and 1024; shorter series (<64) may not be recognized correctly. vLLM support is experimental and may not be stable. Evaluation requires OpenAI API keys for RAGAS.