Salesforce/GiftEvalCTX

GIFT-Eval-CTX

We present GIFT-Eval-CTX, a benchmark for evaluating whether forecasting methods can convert verified future information into accurate numerical trajectories. GIFT-Eval-CTX contains 10,000 instances drawn from 20 GIFT-Eval datasets, covering 7 domains and 6 frequencies, with prediction horizons ranging from 30 to 900 steps. Each instance provides structured effects derived from the realized future: Macro effects describe horizon-wide behavior, while Micro effects specify localized changes.

Useful links: 📄 Paper, 📔 Blog, 🖥️ Code, and 🏆 Leaderboard.

Downloading Data

To load all benchmark instances:

from datasets import load_dataset

dataset = load_dataset("Salesforce/GiftEvalCTX", split="test")

To download the complete repository, including the benchmark data and all methods' results:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Salesforce/GiftEvalCTX",
    repo_type="dataset",
    local_dir="./GiftEvalCTX",
)

Submitting Results

Each method has its own directory under eval/. For example, Moirai-2 results are stored in eval/moirai_2/moirai_2.parquet. All evaluation parquet files contain six fields: dataset, term, item_id, window_id, forecast, and mase. Results are matched to the benchmark data by the four sample-key fields: dataset, term, item_id, and window_id.

Each method directory also contains a config.json with its ID, display name, category, and input information:

{
  "id": "moirai_2",
  "name": "Moirai-2",
  "category": "TSFM",
  "input": "History"
}

To submit a new method, add a folder under eval/ with exactly one Parquet result file and a config.json. We recommend matching the folder and Parquet names to the method ID for clarity. The configuration fields id, name, category, and input must all be nonempty strings, and id must be unique. The prediction results must cover the same sample keys as the benchmark dataset and use the six-field evaluation schema above.

Ethical Considerations

This release is for research purposes only in support of an academic paper. Our models, datasets, and code are not specifically designed or evaluated for all downstream purposes. We strongly recommend users evaluate and address potential concerns related to accuracy, safety, and fairness before deploying this model. We encourage users to consider the common limitations of AI, comply with applicable laws, and leverage best practices when selecting use cases, particularly for high-risk scenarios where errors or misuse could significantly impact people’s lives, rights, or safety. For further guidance on use cases, refer to our AUP and AI AUP.