SciTS: Scientific Time Series Understanding and Generation with LLMs (TimeOmni)

🎉 Accepted to ICLR 2026

📢 The datasets and baseline evaluation scripts will be open-sourced soon.

This is the official SciTS (TimeOmni) repository.

SciTS is a large-scale benchmark for scientific time series understanding and generation across 12 domains and 43 tasks. TimeOmni is a unified framework that equips LLMs with the ability to understand and generate time series while staying compatible with general-purpose LLM training.

---

✨ Highlights

• Unified time series modeling across forecasting, classification, anomaly detection, QA, and more.
• LLM-compatible training pipeline with reprogramming layers.
• Benchmark-ready evaluation with standardized JSONL formats.
• Multimodal time series support (audio, CSV, NumPy, EEG/MEG .fif).

---

🖼️ Figures

Figure 1. SciTS benchmark overview.

Figure 2. TimeOmni framework.

---

📦 Repository Structure

TimeOmni/
├── run_main_refactored_unified.py   # Main training entry
├── infer_benchmark.py               # Distributed inference for JSONL files
├── eval_benchmark.py                # Metric evaluation & aggregation
├── models/                          # TimeOmni model definition
├── layers/                          # Embedding & normalization modules
├── data_provider/                   # Dataset + dataloader factory
├── scripts/                         # Training & inference scripts
├── utils/                           # Helper utilities (early stop, lr schedule, etc.)
├── dataset/                          # Place datasets here (JSONL + raw signals)
├── pretrained/                       # Place pretrained LLM weights here
├── figures/                          # Paper figures
└── ds_config_zero2.json              # DeepSpeed ZeRO-2 config

---

🚀 Quick Start

1) Environment

Use python 3.11 from MiniConda. Install dependencies:

pip install -r requirements.txt

2) Prepare Data

• Place JSONL datasets under dataset/.
• Each JSONL line should contain fields required by Dataset_Unified (see data_provider/dataset.py).
• For raw signals, use supported formats: .wav, .mp3, .flac, .m4a, .csv, .npy, .fif.

3) Pretrained LLM

Put the pretrained weights under pretrained/. Default configuration expects:

pretrained/Qwen3-8B/

4) Train

Use the provided multi-node training script or call the main training entry directly:

bash scripts/TimeOmni_unified.sh

5) Inference & Evaluation

bash scripts/TimeOmni_infer_eval.sh

---

🧠 Core Components

• Training: run_main_refactored_unified.py (Accelerate + DeepSpeed)
• Model: models/TimeOmni.py
• Dataset: data_provider/dataset.py and data_provider/data_factory_unified.py
• Benchmark Inference: infer_benchmark.py
• Benchmark Evaluation: eval_benchmark.py

---

⚙️ Configuration

• DeepSpeed: ds_config_zero2.json
• Training hyperparameters are specified in scripts/TimeOmni_unified.sh.
• Inference & evaluation are configured in scripts/TimeOmni_infer_eval*.sh.

---

📈 Output Artifacts

Training and evaluation outputs are saved under exp/, including:

• config.json (experiment configuration)
• training_log.txt
• infer_results/ (JSONL outputs)
• eval_results/ (CSV metrics)

---

🙏 Acknowledgement

This codebase is adapted from KimMeen/Time-LLM. Thanks to the authors for their excellent work.

---

📜 Citation

If you find this work useful, please cite:

@inproceedings{wu2025scits,
    title={{SciTS: Scientific Time Series Understanding and Generation with LLMs}}, 
    author={Wu, Wen and Zhang, Ziyang and Liu, Liwei and Xu, Xuenan and Liu, Junlin and Fan, Ke and Lv, Qitan and Zhuang, Jimin and Zhang, Chen and Yuan, Zheqi and others},
    booktitle={Proc. ICLR},
    year={2026},
    address={Rio de Janeiro}
}

---

📄 License

This project is licensed under the Apache 2.0 License. See LICENSE for details.