| --- |
| pretty_name: AVeriTeC repaired dev, SPLADE index |
| 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 |
| - splade |
| size_categories: |
| - 10M<n<100M |
| extra_gated_prompt: >- |
| This index is derived from 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 repaired dev, SPLADE index |
|
|
| A prebuilt SPLADE index over |
| [Factiverse/averitec-repaired-dev](https://huggingface.co/datasets/Factiverse/averitec-repaired-dev), |
| so the corpus can be searched without a GPU. |
|
|
| Encoder: `naver/splade-cocondenser-ensembledistil`, the model the |
| AVeriTeC shared task used, so scores are comparable with its baselines. |
|
|
| ## Two differences from the shipped store |
|
|
| **It indexes the repaired text**, in which 63.9% of the empty documents |
| have been recovered (21.5% of documents empty, down to 7.8%). |
|
|
| **It does not drop text.** The shared task's builder flushes its buffer |
| *before* appending the current sentence and never flushes after the |
| loop, so every document loses its final sentence and a single-sentence |
| document yields one empty chunk and no text. This index chunks at 600 |
| words instead, losing nothing. Chunk size was chosen from a sweep over |
| 59 claims where 600 and 300 words gave identical gold recall — paired |
| difference 0.0000, 3 wins, 3 losses, 53 ties — at half the chunks. |
|
|
| ## Layout |
|
|
| ``` |
| dev/full/<claim_id>/documents.pkl chunks, with url metadata |
| dev/full/<claim_id>/embeddings.npz scipy CSR, (n_chunks, 30522) |
| ``` |
|
|
| Row *i* of the matrix is chunk *i* of the pickle. |
|
|
| ```python |
| import pickle, scipy.sparse as sp |
| docs = pickle.load(open("dev/full/2/documents.pkl", "rb")) |
| emb = sp.load_npz("dev/full/2/embeddings.npz") |
| scores = emb @ query_vector # query_vector: SPLADE-encoded, (30522,) |
| ``` |
|
|
| ## Reproducing it |
|
|
| The index is derived, not primary. It is published to save ~3 GPU-hours |
| and to pin the exact chunker, but everything needed to rebuild it is |
| public: the repaired corpus above and |
| `scripts/build_repaired_splade_store.py` from the BEACON repository. |
|
|
| ## Provenance |
|
|
| URLs, document lists and structure come from the |
| [AVeriTeC](https://fever.ai/dataset/averitec.html) shared task, whose |
| terms govern that material. The indexed text was fetched from |
| publishers' own pages and ordinary copyright applies to it. Access is |
| gated for that reason. Please cite AVeriTeC. |
|
|
