DR-MARL paper PDFs — PRIVATE ARCHIVE, DO NOT MAKE PUBLIC
Full-text PDFs for the papers indexed in Ngseo/dr-marl-papers (distributionally-robust RL and cooperative MARL at ICML / NeurIPS / ICLR / AAMAS / AISTATS).
1,841 of 1,989 PDFs · 4.89 GB · average 2.7 MB
⚠️ This repository must stay private
The PDFs are the publishers' copyrighted works, retrieved from their own open-access pages. Keeping a personal copy is one thing; publishing them is another. In particular:
- AAMAS / IFAAMAS: "To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee."
- PMLR (ICML, AISTATS): "Copyright © The authors and PMLR" — per-paper licences vary.
- NeurIPS: no licence notice on the proceedings pages; authors retain copyright.
- OpenReview / ICLR: authors keep copyright.
Flipping this repo to public, or adding collaborators outside the rights you hold, turns a personal archive into redistribution. Don't.
Buckets
pdf_index.jsonl carries the same bucket field as the public index: A robust RL,
B cooperative MARL, C the intersection, D adjacent. The full labelling criteria —
what goes in each bucket, what deliberately does not, and worked examples — are documented in
Ngseo/dr-marl-papers.
| bucket | PDFs | of |
|---|---|---|
| A | 154 | 171 |
| B | 1008 | 1083 |
| C | 60 | 70 |
| D | 619 | 665 |
Layout
{venue}/{year}/{paper_id}.pdf e.g. ICML/2024/ICML_2024_a1b2c3d4e5f6.pdf
pdf_index.jsonl paper_id -> path, route, bytes, title, bucket
paper_id joins directly against papers.csv in the public index repo.
Coverage
| venue | PDFs | of |
|---|---|---|
| AISTATS | 71 | 71 (100%) |
| AAMAS | 673 | 674 (99.9%) |
| NeurIPS | 500 | 521 (96.0%) |
| ICML | 374 | 414 (90.3%) |
| ICLR | 223 | 309 (72.2%) |
By bucket: A 154/171, B 1008/1083, C 60/70, D 619/665.
The 148 missing
All are papers whose only host is OpenReview, which returns 403 to non-browser clients
(ICLR 2018–2026: 86, ICML 2026: 40, NeurIPS 2025: 21, plus one broken AAMAS link). The
arXiv fallback resolved only 14 of the last 147 attempts — recent ICLR/ICML papers often
have no arXiv copy under a matching title. route == "missing" in pdf_index.jsonl marks them.
How these were fetched
src/fetch_pdfs.py in the public index repo. Three routes: the publisher's direct link,
NeurIPS URLs derived from the abstract page, and an open-access/arXiv fallback resolved
through Semantic Scholar. Resumable — it skips files already on disk.
- Downloads last month
- 3