Fix paper link — model is indexed under the WMT19 MT paper instead of BlenderBot
The paper link in this card points to arxiv.org/abs/1907.06616, which is Facebook
FAIR's WMT19 News Translation Task Submission — a machine translation paper. The
intended one is 2004.13637, Recipes for building an open-domain chatbot.
The card already disagrees with itself on that line: the anchor text names the right
paper, the URL points at a different one.
+ Paper: [Recipes for building an open-domain chatbot](https://arxiv.org/abs/1907.06616)
└──────── correct title ────────┘ └──── unrelated paper ────┘
Why this is worth fixing beyond the broken link
There is no arxiv: tag in this card's YAML — the Hub derives it by parsing the arXiv
link out of the card body. So the wrong URL becomes indexed metadata, and the model shows
up on the wrong paper page.
huggingface.co/papers/1907.06616 (the WMT19 MT paper) currently lists 15 models, and
BlenderBot checkpoints are the bulk of them:
| model tagged to the WMT19 paper | downloads | correct? |
|---|---|---|
facebook/blenderbot_small-90M |
97,351 | no |
facebook/blenderbot-3B |
91,686 | no |
facebook/wmt19-ru-en |
15,568 | yes |
facebook/wmt19-en-ru |
6,967 | yes |
facebook/wmt19-en-de |
4,521 | yes |
facebook/wmt19-de-en |
4,099 | yes |
facebook/blenderbot-90M |
725 | no |
facebook/blenderbot-1B-distill |
78 | no |
hyunwoongko/blenderbot-9B |
13 | no (copied card) |
BlenderBot accounts for roughly 190k of the ~221k downloads on that paper's page, so
the four models the paper actually describes are outnumbered on their own page. The real
BlenderBot paper, 2004.13637, lists only 7 models totalling ~7.8k downloads.
To be clear about scope: this doesn't affect academic citation counts, which come from
reference lists in papers. It's the Hub's paper↔model index that's wrong, which is what
drives the paper page and model discovery.
Verification
Both IDs, resolved through the arXiv API:
| arXiv ID | Title | Authors | Published |
|---|---|---|---|
1907.06616 (current) |
Facebook FAIR's WMT19 News Translation Task Submission | Ng, Yee, Baevski, Ott… | 2019-07-15 |
2004.13637 (proposed) |
Recipes for building an open-domain chatbot | Roller, Dinan, Goyal, Ju… | 2020-04-28 |
Two more things that corroborate it, both inside this repo:
- A sibling already has it right.
facebook/blenderbot-400M-distilluses the same
card template with the same anchor text and links2004.13637. This change just brings
this repo in line with it. - The surrounding content agrees. The card also links
parl.ai/projects/recipes/(the
ParlAI Recipes project), and the abstract quoted here opens "Building open-domain
chatbots is a challenging area…" with no mention of translation.
The change
One line, one URL: 1907.06616 → 2004.13637. No code, config or weights touched. Since
the tag is derived from the card body, this should also correct the paper association.
Reported in #1
(Feb 2023). The same link appears on blenderbot-3B, blenderbot-90M andblenderbot-1B-distill — happy to open the same fix there if this looks right.
Hi @lysandre — mind a quick look? One-line card fix.
The paper link here points to arxiv 1907.06616 (WMT19 news translation) rather than 2004.13637 (Recipes for building an open-domain chatbot).
There's no arxiv: tag in the YAML, so the Hub derives the paper link from this URL in the card body — which means the model gets indexed against the wrong paper. Right now huggingface.co/papers/1907.06616 lists 15 linked models, and BlenderBot checkpoints are the top three, above the wmt19-* models the paper is actually about. 4 of the 5 official BlenderBot repos are attributed there; Recipes gets 1.
The line contradicts itself too: the anchor text says "Recipes for building an open-domain chatbot", the URL doesn't. blenderbot-400M-distill already links the right one, so this just matches its sibling.
Resolves #1 (open since Feb 2023). Same link on blenderbot-3B, -90M and -1B-distill if this looks right.