Update LICENSE
Browse files
LICENSE
CHANGED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MS MARCO License Terms
|
| 2 |
+
|
| 3 |
+
The MS MARCO datasets were released by Microsoft in 2016 (initial release) and
|
| 4 |
+
have been updated through several follow-on competitions and BEIR-style
|
| 5 |
+
re-releases. They are distributed under a non-standard set of "MS MARCO" terms
|
| 6 |
+
authored by Microsoft Research — not under a Creative Commons or open-source
|
| 7 |
+
license — granting a non-exclusive, royalty-free right to use the dataset for
|
| 8 |
+
non-commercial research purposes, with attribution.
|
| 9 |
+
|
| 10 |
+
## What you can do
|
| 11 |
+
|
| 12 |
+
- Use the dataset for research, experimentation, and benchmarking.
|
| 13 |
+
- Train and evaluate models on it (passage retrievers, rerankers, generative
|
| 14 |
+
QA, embedding models).
|
| 15 |
+
- Publish numbers against the standard MS MARCO splits (dev / eval / TREC
|
| 16 |
+
Deep Learning subsets) without separate clearance.
|
| 17 |
+
- Redistribute the dataset in unmodified form for research use, with
|
| 18 |
+
attribution back to Microsoft.
|
| 19 |
+
|
| 20 |
+
## What you should still do
|
| 21 |
+
|
| 22 |
+
- **Attribute Microsoft** as the source whenever the dataset is used,
|
| 23 |
+
redistributed, or referenced — including paper citations, blog posts, and
|
| 24 |
+
downstream dataset cards.
|
| 25 |
+
- **Treat commercial deployment as a separate question.** The release terms
|
| 26 |
+
are explicitly research-oriented. Productionising a model that was trained
|
| 27 |
+
on MS MARCO (or evaluating it as part of a paid product) is not addressed
|
| 28 |
+
by the dataset license and should be reviewed independently.
|
| 29 |
+
- **Do not redistribute the held-out evaluation labels.** The TREC Deep
|
| 30 |
+
Learning evaluation qrels are released asynchronously and treating them
|
| 31 |
+
as part of the public corpus undermines the benchmark.
|
| 32 |
+
- **Respect Bing's terms.** The passages are extracted from real Bing web
|
| 33 |
+
search results; redistributing the source URLs or attempting to re-crawl
|
| 34 |
+
the linked pages remains subject to Bing's and the linked sites'
|
| 35 |
+
individual terms.
|
| 36 |
+
|
| 37 |
+
## Background
|
| 38 |
+
|
| 39 |
+
- Original release: Bajaj, Nguyen, et al. — *MS MARCO: A Human Generated
|
| 40 |
+
Machine Reading Comprehension Dataset* (2016).
|
| 41 |
+
- Project site:
|
| 42 |
+
[microsoft.github.io/msmarco/](https://microsoft.github.io/msmarco/)
|
| 43 |
+
- Passage retrieval task overview:
|
| 44 |
+
[microsoft.github.io/msmarco/Datasets#passage-ranking-dataset](https://microsoft.github.io/msmarco/Datasets#passage-ranking-dataset)
|
| 45 |
+
|
| 46 |
+
## SPDX identifier
|
| 47 |
+
|
| 48 |
+
No formal SPDX identifier is registered. Catalogued here as
|
| 49 |
+
`ms-marco-terms`.
|