xiezhe24/ChatTS-14B-GPTQ-Int4

[VLDB' 25] ChatTS-14B-GPTQ-4Bit Model

This is the GPTQ-4Bit quantized model of ChatTS-14B.

[VLDB' 25] ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning

ChatTS focuses on Understanding and Reasoning about time series, much like what vision/video/audio-MLLMs do. This repo provides code, datasets and model for ChatTS: ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning.

Key Features

ChatTS is a Multimodal LLM built natively for time series as a core modality：

• ✅ Native support for multivariate time series
• ✅ Flexible input: Supports multivariate time series with different lengths and flexible dimensionality
• ✅ Conversational understanding + reasoning:
  Enables interactive dialogue over time series to explore insights about time series
• ✅ Preserves raw numerical values:
  Can answer statistical questions, such as "How large is the spike at timestamp t?"
• ✅ Easy integration with existing LLM pipelines, including support for vLLM.

Example Application

Here is an example of a ChatTS application, which allows users to interact with a LLM to understand and reason about time series data:

Link to the paper

Link to the Github repository

Usage

• This model is fine-tuned on the QWen2.5-14B-Instruct (https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) model. For more usage details, please refer to the README.md in the ChatTS repository.
• An example usage of ChatTS (with HuggingFace):

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
import torch
import numpy as np

hf_model = "bytedance-research/ChatTS-14B"
# Load the model, tokenizer and processor
# For pre-Ampere GPUs (like V100) use `_attn_implementation='eager'`
model = AutoModelForCausalLM.from_pretrained(hf_model, trust_remote_code=True, device_map="auto", torch_dtype='float16')
tokenizer = AutoTokenizer.from_pretrained(hf_model, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(hf_model, trust_remote_code=True, tokenizer=tokenizer)
# Create time series and prompts
timeseries = np.sin(np.arange(256) / 10) * 5.0
timeseries[100:] -= 10.0
prompt = f"I have a time series length of 256: <ts><ts/>. Please analyze the local changes in this time series."
# Apply Chat Template
prompt = f"""<|im_start|>system
You are a helpful assistant.<|im_end|><|im_start|>user
{prompt}<|im_end|><|im_start|>assistant
"""
# Convert to tensor
inputs = processor(text=[prompt], timeseries=[timeseries], padding=True, return_tensors="pt")
# Model Generate
outputs = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(outputs[0][len(inputs['input_ids'][0]):], skip_special_tokens=True))

Reference

• QWen2.5-14B-Instruct (https://huggingface.co/Qwen/Qwen2.5-14B-Instruct)
• transformers (https://github.com/huggingface/transformers.git)
• ChatTS Paper

License

This model is licensed under the Apache License 2.0.

Cite

@article{xie2024chatts,
  title={ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning},
  author={Xie, Zhe and Li, Zeyan and He, Xiao and Xu, Longlong and Wen, Xidao and Zhang, Tieying and Chen, Jianjun and Shi, Rui and Pei, Dan},
  journal={arXiv preprint arXiv:2412.03104},
  year={2024}
}