| --- |
| language: |
| - en |
| task_categories: |
| - question-answering |
| tags: |
| - frames |
| - wikipedia |
| - multi-hop-qa |
| - offline-corpus |
| license: other |
| license_name: mixed-apache-2.0-cc-by-sa-4.0 |
| license_link: https://huggingface.co/datasets/zhs039/frames-oracle-corpus/blob/main/README.md#licenses-and-attribution |
| --- |
| |
| # FRAMES Oracle Corpus |
|
|
| An offline Wikipedia document corpus derived from [google/frames-benchmark](https://huggingface.co/datasets/google/frames-benchmark), packaging each question with its ground-truth reference articles in Markdown. |
|
|
| ## Contents |
|
|
| All dataset files are located under `corpus/`: |
|
|
| - `corpus/frames.tsv`: 822 questions and reference answers from FRAMES, retaining original row IDs and columns. |
| - `corpus/index.json`: Maps `questions` (question IDs to document IDs) and `documents` (document IDs to title, source URL, and Markdown filename). |
| - `corpus/doc_0001.md` – `corpus/doc_2500.md`: 2,500 extracted Wikipedia articles in Markdown. |
|
|
| ## Preparation & Notes |
|
|
| - **Source revision**: Built from `google/frames-benchmark` (`test.tsv` at commit `58d9fb6330f3ab1316d1eca12e5e8ef23dcc22ef`). |
| - **Filtering & link resolution**: Questions 141 and 540 are excluded due to missing English biographies. Question 823 retains its original Simple English Wikipedia article, and Question 650 retains its category-page content. Combined URL annotations were split, and malformed URLs in questions 644 and 764 were resolved to their target English pages. |
| - **Formatting**: Wikipedia HTML was converted to clean article-body Markdown, preserving tables while stripping navigation, scripts, and reference clutter. Note that these are static snapshots downloaded after benchmark publication rather than historical Wikipedia revisions. |
|
|
| ## Download |
|
|
| ```bash |
| uvx --from huggingface_hub hf download zhs039/frames-oracle-corpus \ |
| --repo-type dataset --include "corpus/*" --local-dir . |
| ``` |
|
|
| ## License & Attribution |
| - Questions and answers (corpus/frames.tsv): Derived from Google's FRAMES benchmark under the Apache License 2.0 (rows 141 and 540 removed; packaged with local document mappings). |
|
|
| - Wikipedia articles (corpus/doc_*.md): Copyright belongs to the respective Wikipedia contributors and is distributed under CC BY-SA 4.0 International. Article titles and source URLs are preserved in corpus/index.json. |
