beacon-data / README.md
facti's picture
Add dataset card
5a37b88 verified
|
Raw History Blame Contribute Delete
2.46 kB
---
license: other
license_name: mixed-see-card
pretty_name: BEACON derived SPLADE indexes
---
# BEACON derived indexes
SPLADE indexes built for *Coverage Is Not the Bottleneck: A Negative
Result for Budgeted Evidence Acquisition in Fact-Checking*. Each holds
chunked passages and their SPLADE cocondenser-ensembledistil vectors in
the layout the code reads: `documents.pkl` and `embeddings.npz`.
Code: <https://github.com/factiverse/beacon>
## Contents
| Directory | Size (GiB) | Source licence | Redistribution note |
|---|---|---|---|
| `raguard` | 0.0 | MIT | RAGuard is released under the MIT licence, which permits redistribution of derived works with attribution. |
| `quantemp` | 0.4 | CC BY-NC 4.0 | QuanTemp is CC BY-NC 4.0. The NonCommercial clause makes redistribution from an organisational account a decision for the dataset owners, not a default. |
| `trec_hm` | 12.5 | TREC / C4 derived | The TREC 2022 Health Misinformation corpus derives from C4 and is governed by the track's data agreement. Document identifiers may be shared; document text may not. |
The AVeriTeC repaired corpus and its index are published separately:
[`Factiverse/averitec-repaired-dev`](https://huggingface.co/datasets/Factiverse/averitec-repaired-dev)
and
[`Factiverse/averitec-repaired-dev-splade`](https://huggingface.co/datasets/Factiverse/averitec-repaired-dev-splade).
## Licence and your obligations
These are **derived works**. The passages are chunks of documents
published by other people, and the vectors are computed from those
chunks. This repository is gated for that reason: by requesting access
you accept that you must comply with each source corpus's own licence,
listed above, and that the terms differ between directories. In
particular, the QuanTemp material is NonCommercial, and the TREC Health
Misinformation material derives from a corpus governed by the track's
data agreement.
If your use does not fit those terms, rebuild the index you need from
the source corpus instead: `scripts/fetch_datasets.py` retrieves what is
publicly available and `scripts/reproduce.sh index <dataset>` rebuilds
the index. The upload exists to save GPU hours, not to widen any
licence.
## Provenance
Every index was built by this repository's own pipeline; nothing was
downloaded pre-built. The AVeriTeC corpus additionally received a repair
pass documented in the paper, since 21.6% of its dev documents and 31.1%
of its test documents ship empty.