GIFT-CTX / README.md
doyensahoo's picture
Initial upload of parquet dataset configs
231fe8b verified
metadata
configs:
  - config_name: gift_ctx
    data_files:
      - split: train
        path: gift_ctx/train.parquet

Gift-EvalCTX Parquet

This repository hosts the GIFT-EvalCTX dataset in parquet format, highly compatible with LLMs. Each row is a sample of the dataset and contain the following fields:

  • idx: unique id of the sample
  • source: source of the same
  • skill: the types of context
  • frequency: frequency of the sample
  • history_values: a string containing values of the history
  • history_start: starting timestamp of the history
  • history_end: ending timestamp of the history
  • future_values: a string containing values of the forecasting future
  • future_start: starting timestamp of the future
  • future_end: ending timestamp of the future
  • entry_sep: separator used to split the string in history_values, future_values, and roi into array of floats
  • roi: region of interest, indicating the indices of the timestamp that are affected by the context and used in evaluation
  • pred_length: number of timestamps to forecast
  • system_prompt: default system prompt
  • user_instruct: query to forecast
  • context_info: context for the current sample
  • prompt: a complete prompt to query LLM

Note that all fields contain string only, you need to convert them into the appropriate format (array of floats, or datetime).

Example usage

from datasets import load_dataset

ds = load_dataset(
    "Salesforce/GiftEvalCTX",
    "gift_ctx",
    split="train"
)
print(len(ds))
print(ds[0].keys())

This repository is made public for research purposes only.