| --- |
| pretty_name: AVeriTeC knowledge store, extraction repaired (dev) |
| language: |
| - en |
| license: other |
| license_name: averitec-derived |
| license_link: https://fever.ai/dataset/averitec.html |
| task_categories: |
| - text-retrieval |
| tags: |
| - fact-checking |
| - evidence-retrieval |
| - averitec |
| - fever |
| size_categories: |
| - 100K<n<1M |
| extra_gated_prompt: >- |
| This corpus contains text extracted from third-party web pages, including |
| news articles. It is released for research on evidence retrieval for |
| fact-checking. By requesting access you agree to use it for research |
| purposes, to respect the rights of the original publishers, and not to |
| redistribute the extracted text. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| --- |
| |
| # AVeriTeC knowledge store, with extraction repaired (dev split) |
|
|
| The AVeriTeC dev knowledge store with its empty documents re-scraped. |
|
|
| In the original store, 21.5% of documents have an empty `url2text`: the |
| scraper ran, got nothing back, and the empty result was indexed anyway. |
| Roughly a third of gold evidence documents were among them, so retrieval |
| results on that store are bounded by a ceiling unrelated to the method. |
|
|
| | | documents | |
| |---|---| |
| | total | 506,640 | |
| | empty before | 109,151 (21.5%) | |
| | empty after | 39,378 (7.8%) | |
| | recovered | **69,773 (63.9%)** | |
|
|
| ## Using it |
|
|
| A drop-in replacement for the original split directory — same file |
| names, same JSON Lines format, same records in the same order. |
|
|
| ``` |
| dev/<claim_id>.json one record per document: claim_id, query, type, |
| url, url2text (a list of sentences) |
| MANIFEST.json per-claim document and empty counts |
| ``` |
|
|
| ```python |
| import json |
| records = [json.loads(line) for line in open("dev/2.json")] |
| ``` |
|
|
| Documents that already had text are byte-identical to the original, so a |
| comparison against the original store differs only by recovered text. |
|
|
| ## Limitations |
|
|
| Recovery tops out near 64%: 39,378 documents are still empty, mostly |
| publishers that block scrapers. Some recoveries are page fragments |
| rather than articles. Repairing text does not by itself make evidence |
| retrievable. |
|
|
| ## Provenance |
|
|
| URLs, document lists and structure come from the |
| [AVeriTeC](https://fever.ai/dataset/averitec.html) shared task, whose |
| terms govern that material. The recovered text was fetched from |
| publishers' own pages and ordinary copyright applies to it, exactly as |
| for the original store. Access is gated for that reason. Please cite |
| AVeriTeC. |
|
|
