Papers — what may be mirrored here, what may not, and which twelve are¶
Every work cited in ../REFERENCE.md and the per-experiment
METHOD.md files, with the redistribution licence read from the publisher's own page
on 2026-09-29 rather than assumed.
fetch_papers.py downloads the redistributable ones into this
directory for offline reading. Twelve of the twenty-four are committed here; the
other twelve are gitignored by name so --all cannot commit them by accident. See
What is committed here.
(This paragraph and the heading above used to say nothing was mirrored, which was true only before the mirroring was done. The body of this file was right and the header was stale — corrected 2026-09-29.)
python docs/papers/fetch_papers.py # the 12 that permit redistribution
python docs/papers/fetch_papers.py --all # all 24, for personal use only
Redistributable — CC BY / CC BY-SA, attribution required¶
Mirroring these is permitted provided the author and licence travel with the copy.
| Work | Licence | Source |
|---|---|---|
| Lin (2004), ROUGE: A Package for Automatic Evaluation of Summaries | CC BY 4.0 | ACL W04-1013 |
| Tjong Kim Sang & De Meulder (2003), Introduction to the CoNLL-2003 Shared Task | CC BY 4.0 | ACL W03-0419 |
| Dror et al. (2018), The Hitchhiker's Guide to Testing Statistical Significance in NLP | CC BY 4.0 | ACL P18-1128 |
| Kryscinski et al. (2019), Neural Text Summarisation: A Critical Evaluation | CC BY 4.0 | ACL D19-1051 |
| Wadden et al. (2020), Fact or Fiction: Verifying Scientific Claims | CC BY 4.0 | ACL 2020.emnlp-main.609 |
| Ding et al. (2021), Few-NERD: A Few-shot Named Entity Recognition Dataset | CC BY 4.0 | ACL 2021.acl-long.248 |
| Reimers & Gurevych (2019), Sentence-BERT | CC BY-SA 4.0 | arXiv 1908.10084 |
| Thakur et al. (2021), BEIR | CC BY-SA 4.0 | arXiv 2104.08663 |
| He et al. (2021), DeBERTaV3 | CC BY 4.0 | arXiv 2111.09543 |
| Wang et al. (2022), Text Embeddings by Weakly-Supervised Contrastive Pre-training (E5) | CC BY 4.0 | arXiv 2212.03533 |
| Zaratiana et al. (2023), GLiNER | CC BY 4.0 | arXiv 2311.08526 |
| Zhang et al. (2024), A Careful Examination of LLM Performance on Grade School Arithmetic | CC BY 4.0 | arXiv 2405.00332 |
NOT redistributable — link only¶
| Work | Licence | Why not |
|---|---|---|
| Hermann et al. (2015), Teaching Machines to Read and Comprehend | arXiv non-exclusive 1.0 | Grants arXiv the right to distribute, not third parties |
| Zhang, Zhao & LeCun (2015), Character-level Convolutional Networks | arXiv non-exclusive 1.0 | same |
| Nallapati et al. (2016), Abstractive Text Summarisation Using Seq2Seq RNNs | arXiv non-exclusive 1.0 | same |
| See, Liu & Manning (2017), Get To The Point | arXiv non-exclusive 1.0 | same |
| Guo et al. (2017), On Calibration of Modern Neural Networks | arXiv non-exclusive 1.0 | same |
| Devlin et al. (2018), BERT | arXiv non-exclusive 1.0 | same |
| Yin et al. (2019), Benchmarking Zero-shot Text Classification | arXiv non-exclusive 1.0 | same |
| Lewis et al. (2019), BART | arXiv non-exclusive 1.0 | same |
| Song et al. (2020), MPNet | arXiv non-exclusive 1.0 | same |
| Xiao et al. (2023), C-Pack (BGE) | arXiv non-exclusive 1.0 | same |
| Järvelin & Kekäläinen (2002), Cumulated gain-based evaluation of IR techniques | © ACM | ACM DL, subscription |
| Holm (1979), A Simple Sequentially Rejective Multiple Test Procedure | © Scandinavian Journal of Statistics | Paywalled. Also the one citation in this repo that could not be verified programmatically — JSTOR serves a bot challenge and Crossref does not index the 1979 volume, so the reference is given in full text instead of linked |
| Demšar (2006), Statistical Comparisons of Classifiers over Multiple Data Sets | not machine-readable on the page | JMLR v7 states no licence this check could read; freely readable there |
The arXiv non-exclusive licence is the common trap. It reads like an open licence and is not: it grants arXiv a perpetual right to distribute the work. It says nothing about anyone else mirroring it. Ten of the sixteen arXiv papers cited here are under it.
What is committed here¶
The twelve CC-licensed papers above are mirrored in this directory, 18 MB, with
ATTRIBUTION.md carrying the author, title, licence and source of each —
which is what CC BY and CC BY-SA require of anyone redistributing them. None is modified;
each is the publisher's PDF byte-for-byte.
The other twelve are not here and are gitignored by name, so --all cannot commit them
by accident.
Two caveats a reader of this directory should have:
- The split follows publisher licensing policy, not importance. Twelve papers sitting here does not make them the canon. nDCG (Järvelin & Kekäläinen) and Holm — two of the most load-bearing citations in this repo — are both in the un-mirrorable half. Read the table above, not the file listing.
- Two are CC BY-SA, which is share-alike: a derivative of them must carry the same licence. Unmodified redistribution, as here, does not trigger that.
This is the one exception to the rule everywhere else in the repo — no corpora, no weights,
no third-party bytes, only recipes. It is an exception the licences explicitly permit, and
the obligation that comes with it is discharged in ATTRIBUTION.md.