InstructTime v2.1

What's New

Cross-Domain Autoregressive Pretraining

• Added run_pretrain_universal.py for multi-domain joint pretraining across all five signal domains (ECG, EEG, FD, HAR, RWC)
• Supports weighted domain sampling for balanced cross-domain training
• Mixed-precision training with cosine warmup scheduling and early stopping
• Deterministic reproducibility setup for consistent results

Bug Fixes

• Fixed off-by-one error in average loss and metric calculation in TStokenizer Trainer class (loss_sum / idx → loss_sum / (idx + 1))

Documentation

• Updated README to reflect the correct three-stage training pipeline:
  1. TStokenizer training
  2. Cross-domain autoregressive pretraining (run_pretrain_universal.py)
  3. Supervised fine-tuning (Universal + Adaptation)
• Added run_pretrain_universal.py to project structure

Changed Files

• run_pretrain_universal.py (new)
• TStokenizer/process.py
• README.md

InstructTime v2.0 – Refactored Codebase

Highlights

• Refactored the entire codebase for better modularity, readability, and extensibility.
• Unified project structure for TStokenizer, InstructTime-Universal, and InstructTime-Adapt.
• Simplified training & evaluation pipelines, with clearer entry scripts and configuration management.
• Improved dataset and preprocessing utilities to make reproducing experiments more robust and user-friendly.
• Updated documentation and examples to match the new code structure.

Migration Notes

• The overall functionality is consistent with v1.x, but file paths, module imports, and script entry points may have changed.
• Please check the updated README and example commands before re-running previous experiments.

InstructTime v1.0 – Initial Release

Highlights

• First stable release of InstructTime for time series classification with multimodal language modeling (ACM WSDM 2025).
• Includes training pipeline for TStokenizer, InstructTime-Universal, and InstructTime-Adapt.
• Provides scripts for data preprocessing, multi-dataset loading, model training, and evaluation (preprocess.py, multidataset.py, multimodel.py, metrics.py, etc.).
• Example prompts and usage instructions for EEG / sleep stage classification are included in the README.
• Intended for researchers and practitioners to reproduce and extend the results reported in the InstructTime paper.

For installation and usage details, please refer to the README.