Data

This directory contains the datasets used for training and evaluation. Due to size constraints, most data must be downloaded externally before use.

Directory Structure

data/
├── README.md
├── glucose/
│   └── y_glucose.csv            # small datasets included in repo
├── ...
├── reviews/
│   └── (see below)
├── gifteval/
│   └── (see below)
└── tsmixup/
    └── (see below)

Setup

Glucose, M5 (Shopping), Meditation, Patents, Police

Included in the repository. No action needed.

Reviews (Amazon-Google)

This dataset is a preprocessed aggregation of Google Local Reviews and Amazon Reviews which is too large to include in the repository. Download the preprocessed csv file from the UCI Machine Learning Repository

GiftEval Subsample

Download from HuggingFace:

huggingface-cli download Salesforce/GiftEval \
    --repo-type=dataset \
    --local-dir data/gifteval

Then run the preprocessing notebook:

data/gifteval/setup_gifteval.ipynb

TSMixup (Training)

The training data used for training prediction heads uses the Chronos TSMixup 10M corpus hosted on HuggingFace. It is downloaded, processed, and cached automatically by src/utils/data_loader.py.

To trigger the download and processing manually:

from src.utils.data_loader import create_cached_tsmixup_datasets

train_dataset, val_dataset = create_cached_tsmixup_datasets(
    max_samples=100000,        # adjust as needed
    context_length=512,
    prediction_length=128,
    cache_dir="data/tsmixup",  # raw HF cache location
    processed_cache_path="data/tsmixup/tsmixup_processed.pkl",
)

This will:

1. Download the raw dataset from HuggingFace (autogluon/chronos_datasets, config training_corpus_tsmixup_10m) into cache_dir.
2. Filter, validate, and preprocess each time series (length/quality checks, NaN/Inf removal, etc.).
3. Save the processed result as a pickle file at processed_cache_path.

Subsequent runs will load directly from the pickle cache unless force_reprocess=True is passed.