Reranker research data and model attribution
===========================================

Only original research code and aggregate measurements are committed here.
Dataset texts, judgments, per-query rankings and model weights are not bundled
in Git or the wheel. This notice does not grant rights to redistribute them.
No CLDR or other new system-entity catalogs were added; their NOTICE is unchanged.

BEIR SciFact / NFCorpus / FiQA
-----------------------------
Credit the SciFact (AllenAI), NFCorpus and FiQA benchmark creators, their
contributors, and the BEIR project (Thakur et al.) for its benchmark packaging.
Sources used:
https://github.com/allenai/scifact
https://github.com/beir-cellar/beir
https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/scifact.zip
https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/nfcorpus.zip
https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/fiqa.zip
The local BEIR archives do not carry dataset licence notices. Underlying
corpus/judgment licences have not been independently verified offline; confirm
upstream terms before redistribution. BEIR's code licence is not a licence
for every dataset. This round adds owner-approved local FiQA evaluation,
not a licensed corpus release. Archive checksums are in the evidence manifest.

C-MTEB/T2Reranking
-----------------
Credit the T2Ranking dataset creators and C-MTEB contributors.
https://huggingface.co/datasets/C-MTEB/T2Reranking
Revision: 76631901a18387f85eaa53e5450019b87ad58ef9
The source card reviewed in the previous round declared no explicit licence.
Licence status: unverified; no corpus redistribution approval is asserted.
The fifty-judged-query corpus is a derived local pool of all source dev-row
positive/negative passages, not the original full T2Ranking retrieval corpus.

Multilingual E5-small
--------------------
Credit intfloat and the E5 model contributors.
https://huggingface.co/intfloat/multilingual-e5-small
Revision: 614241f622f53c4eeff9890bdc4f31cfecc418b3
Model card declares MIT. Weights remain in the shared model cache, not Git.
Retain upstream MIT copyright/permission notices if redistributing weights.
Other tools are normal research dependencies, not copied third-party catalogs.

Mr. TyDi
--------
Credit the Mr. TyDi and TyDi QA creators (Castorini / Google Research and
contributors). https://huggingface.co/datasets/castorini/mr-tydi
Topics revision 1d43c80218d06d0ef80f5b172ccabd848b948bc1;
corpus revision 3a3aa212bbe94a8cc0dc858710a3dad49d532054.
Local banks are 30 test queries plus judged documents and 30,000 seeded extra
passages, not a full Wikipedia dump and not a corpus release.
