README.md

metadata

license: apache-2.0
task_categories:
  - question-answering
language:
  - en

📚 mmrag benchmark

📁 Files Overview

• mmrag_train.json: Training set for model training.
• mmrag_dev.json: Validation set for hyperparameter tuning and development.
• mmrag_test.json: Test set for evaluation.
• processed_documents.json: The document pool used for document retrieval.

---

🔍 Query Datasets: mmrag_train.json, mmrag_dev.json, mmrag_test.json

The three files are all lists of dictionaries. Each dictionary contains the following fields:

🔑 id

• Description: Unique query identifier, structured as dataset_queryID.
• Example: ott_144.

❓ query

• Description: Text of the question.
• Example: "What is the capital of France?"

✅ answer

• Description: The gold-standard answer corresponding to the query.
• Example: "Paris"

📑 relevant_chunks

• Description: Dictionary of relevant document IDs and their corresponding relevance scores. The document IDs are structured as dataset_documentID_chunkIndex, equivalent to dataset_queryID_chunkIndex
• Example: {"ott_23573_2": 1}

📖 ori_context

• Description: A list of the original document IDs related to the query.
• Example: ["ott_6104"]

📜 dataset_score

• Description: The relevance score of all datasets regarding this query.
• Example: {"tat": 0, "triviaqa": 2, "ott": 4, "kg": 1, "nq": 0}

---

📚 Knowledge Base: processed_documents.json

This file is a list of documents used for document retrieval, which contains the following fields:

🔑 id

• Description: Unique document identifier, structured as dataset_documentID_chunkIndex, equivalent to dataset_queryID_chunkIndex
• example: ott_8075_0

📄 text

• Description: Text of the document.
• Example: A molecule editor is a computer program for creating and modifying representations of chemical structures.

---

📄 License

This dataset is licensed under the Apache License 2.0.