Weather Captioned – First Time Series – Text Multimodal Dataset

Data Sources

• Time‑series data come from the Max‑Planck‑Institute for Biogeochemistry, Jena WS Beutenberg station.
• Weather forecast reports are obtained from publicly available forecasting platforms.

About the Descriptions

• Descriptions are generated from raw data sourced from public weather forecast services.
• No time‑series data were provided to the large language model.
• Descriptions were generated by GPT‑4; the generation script is located at data_process_scripts/weather_caption.py.
• The total cost of generating the dataset descriptions may exceed $400.
• Users are encouraged to generate their own descriptions.

About the Pre‑Embeddings

• Pre‑embedded news texts are provided.
• Embeddings can be downloaded here.
• The embedding generation scripts are located at data_process_scripts/embedding_caption_local.ipynb and data_process_scripts/embedding_caption.ipynb.

Data Processing Workflow

• Two hash tables manage news data to align it chronologically with the time series.
• News text embeddings are saved as .npy files, with file names serving as hash keys.
• Time‑stamped series segments are converted to a list of news hash keys via the Date2Hash table, then embeddings are read via the Hash2Emb table.
• The Hash2Text table can be used to inspect the news text corresponding to a hash key.