# WattBot 2025 — Data

来源：https://www.kaggle.com/competitions/WattBot2025/data

采集时间：2026-10-20T09:04:53Z

# Dataset Description

This dataset provides metadata and question files used in the WattBot challenge. Participants must build retrieval-augmented generation (RAG) systems that reason over these materials to answer questions about the energy use and environmental impact of different AI workloads.

Participants are responsible for downloading, parsing, chunking, embedding, and retrieving from these documents. No preprocessed text is provided.

## Files

- **`metadata.csv`**  
  A machine-readable index of all referenced documents with the following fields:

  - `id`: unique identifier for the document (e.g., `strubell2019`)  
  - `type`: document category (`paper`, `report`, etc.)  
  - `title`: document title  
  - `year`: publication year  
  - `citation`: formatted citation string  
  - `url`: link to the original PDF or webpage  

- **`train_QA.csv`**  
  A set of annotated question–answer pairs to help you calibrate your RAG pipeline. The schema matches the solutions file.

  Fields include:  
  - `id`: question ID  
  - `question`: full question text  
  - `answer`: natural language response (fallback phrase if unanswerable)  
  - `answer_value`: normalized numeric or categorical value (or `is_blank`)  
  - `answer_unit`: unit of measurement (or `is_blank`)  
  - `ref_id`: list of supporting document IDs from `metadata.csv` (or `is_blank`)  
  - `ref_url`: URL(s) of the cited document(s) (or `is_blank`)  
  - `supporting_materials`: verbatim evidence such as a quote, table/figure reference, or figure caption (or `is_blank`)  
  - `explanation`: reasoning or calculation steps linking the evidence to the answer (or `is_blank`)

- **`test_Q.csv`**  
A larger set of questions with ground-truth answers hidden. You must generate predictions for these questions using your RAG system. Formatting conventions must match `train_QA.csv`