Hugging Face Paper Pages: Daily Papers, arXiv Links, and Model Metadata

A Hugging Face paper page lives at huggingface.co/papers/<arXiv-ID> (for example /papers/2310.06825, the Mistral 7B paper). It mirrors the arXiv abstract and adds upvotes, comments, verified authors, and auto-built lists of the models, datasets and Spaces that cite the paper. huggingface.co/papers itself is Daily Papers, a curated submission feed.

Two other things people mean by this query: the papers Hugging Face’s own team published, and the metadata plumbing that connects a paper to downloadable weights. Both are covered below.

What is actually on a paper page

Section	What it is	Who controls it
Abstract	Mirrored text, plus a link to the arXiv PDF	arXiv
Upvotes	A popularity signal. Not review, not replication	Any logged-in account
Authors	Author names, linked to Hub profiles once authorship is claimed and confirmed	Hugging Face verifies claims
Models / Datasets / Spaces citing this paper	Derived automatically from repo cards that reference the arXiv ID	Repo owners
Community tab	Threaded comments, often including author replies	Any logged-in account

In practice a paper page exists once the paper has been submitted to Daily Papers or referenced from a Hub repository. If you follow an arXiv link and land on a 404, nobody has connected it to the Hub yet.

Daily Papers versus paper pages

They are different objects. The paper page is permanent and keyed on the arXiv ID. Daily Papers is the dated feed at huggingface.co/papers, where community submissions are collected per day and ordered largely by upvotes.

Submission is gated: you generally need to be a verified author of the paper to put it in the feed, and there is a cap on how many you can submit. The exact cap and cutoff time have changed over the life of the feature, so read the notice on the submission form rather than trusting a number you found in a blog post.

URL patterns worth knowing

URL	What you get
huggingface.co/papers	Today’s Daily Papers
huggingface.co/papers?date=2026-03-04	That specific day’s feed
huggingface.co/papers/2310.06825	The paper page for that arXiv ID
huggingface.co/papers?q=speculative+decoding	Search across indexed papers

The date and search parameters are stable enough to bookmark. Tab names on the index (trending, weekly, monthly views) have been reshuffled more than once, so navigate those from the page itself.

How models get linked to a paper: README.md and model-index

Every repository on the Hub — model, dataset or Space — has a README.md at its root. That file is the card. It opens with a YAML block fenced by two --- lines, which the Hub parses for structured metadata, and everything after the closing fence renders as Markdown.

Getting into “Models citing this paper”

Two mechanisms, and you can use both:

• Put the plain arXiv URL (https://arxiv.org/abs/2310.06825) in the body of README.md, usually in the citation block. The Hub extracts arXiv links from cards.
• Add the explicit tag arxiv:2310.06825 to the tags list in the YAML front matter. This is the unambiguous form, and it makes the repo filterable by that tag.

Reporting the paper’s numbers with model-index

model-index is a YAML key in the same front matter. It is how self-reported evaluation results become machine-readable instead of a Markdown table nobody can query. It is an array of models, each with a list of results, and each result binds a task, a dataset and one or more metrics:

---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
base_model: mistralai/Mistral-7B-v0.3
datasets:
- squad
tags:
- arxiv:2310.06825
model-index:
- name: my-paper-reproduction
  results:
  - task:
      type: question-answering
      name: Question Answering
    dataset:
      type: squad
      name: SQuAD v1.1
      split: validation
    metrics:
    - type: exact_match
      value: 87.3
      name: Exact Match
      verified: false
---

Three details that trip people up. The key is hyphenated (model-index, not model_index). task.type must be a valid Hub task identifier such as text-generation or question-answering, otherwise the block is ignored rather than rejected loudly. And verified: false is the honest default — verified results come from Hugging Face’s own evaluation runs, not from you asserting them.

If your YAML is malformed the card still renders; the metadata just silently vanishes. Check the repo page after committing, and compare your card’s rendered specs against a reference such as the Convly models database to confirm the fields you expect actually landed.

Reproducing a paper’s model locally

Before downloading 30 GB of weights, check whether the model fits your card. Run the number through the VRAM calculator, or look it up in the VRAM requirements table by model. A paper’s headline result is often produced on 8×H100; the checkpoint on the paper page may be a much smaller distillation.

The tooling is one Python package on all three platforms:

pip install -U huggingface_hub
hf auth login
hf download mistralai/Mistral-7B-Instruct-v0.3 --include "*.safetensors" "*.json" "tokenizer*"

The CLI was renamed from huggingface-cli to hf in a 2025 release of huggingface_hub. Older installs use huggingface-cli download and huggingface-cli login; if hf is not found, that is why. Run hf --help or pip show huggingface_hub to see which you have.

Linux

Cache defaults to ~/.cache/huggingface/hub. Relocate it with export HF_HOME=/mnt/models/hf before running anything. For large repos, pip install hf_transfer and export HF_HUB_ENABLE_HF_TRANSFER=1 raises throughput substantially on fast connections. This is the path where custom CUDA kernels shipped with a paper’s code will actually build; see GPU picks for local LLMs if you are choosing hardware for reproduction work.

macOS

Same cache path, ~/.cache/huggingface/hub. On Apple Silicon, PyTorch’s MPS backend covers ordinary transformer inference, but paper repos that depend on hand-written CUDA extensions will not compile at all. When that happens, the pragmatic route is a quantized GGUF conversion through Ollama — you lose exact numerical parity with the paper, so do not report it as a reproduction.

Windows

Cache defaults to C:Users<you>.cachehuggingfacehub. Set a different location in PowerShell with $env:HF_HOME="D:hf" for the session, or through System Properties for something permanent. Without Developer Mode or an elevated shell, the Hub cache cannot create symlinks and falls back to duplicating files, which costs extra disk and prints a warning on every download; HF_HUB_DISABLE_SYMLINKS_WARNING=1 quiets the message without changing the behaviour. Many research repos assume a Unix shell, so WSL2 is usually less friction than native Windows for anything past inference.

Platform	Default cache	Main gotcha
Linux	~/.cache/huggingface/hub	None; the reference environment
macOS	~/.cache/huggingface/hub	No CUDA; custom kernels won’t build
Windows	C:Users<you>.cachehuggingfacehub	Symlink fallback duplicates files

Papers published by Hugging Face itself

The two most cited are easy to pin down:

Paper	arXiv ID
Transformers: State-of-the-Art Natural Language Processing	1910.03771
DistilBERT, a distilled version of BERT	1910.01108

Later work — the Zephyr alignment papers, the FineWeb dataset paper, the SmolLM series — is best located from the source rather than from a remembered ID. Open the relevant org profile (huggingface.co/HuggingFaceTB, for instance) and use its Papers tab, or open a model card and read its citation block. Both routes give you the current canonical reference; a hardcoded list in an article does not.

Pulling paper data programmatically

The stable, documented route is the ordinary Hub API, which works because arxiv:<id> is a normal tag:

curl "https://huggingface.co/api/models?filter=arxiv:2310.06825"

That returns every model repo declaring that paper. There is also a JSON endpoint behind the Daily Papers feed itself, but it is undocumented and has no compatibility guarantee — fine for a personal digest script, not something to build a product on. If you need durable paper metadata, scrape from arXiv’s official API and join on the ID.

Frequently asked questions

Is a Hugging Face paper page the same as the arXiv page?

No. It mirrors the abstract and links out to the arXiv PDF, but it is a separate object with its own upvotes, comments and repo links. For citations, always use the arXiv entry or the published venue version. Referencing the Hugging Face URL in a bibliography is a common mistake in preprints and reviewers notice.

How do I get my model to show up under “Models citing this paper”?

Add the arXiv link to your repo’s README.md, or add arxiv:<id> to the tags list in its YAML front matter. The listing is generated from card metadata, so it updates after you commit the card, not when you upload weights. If it does not appear, the usual cause is broken YAML in the front matter.

Can I submit someone else’s paper to Daily Papers?

Generally no — submission is restricted to verified authors of the paper. You can still upvote, comment, and share an existing page, and linking the arXiv ID from a repo you own is enough to get a paper page to exist. Authorship claims are made from the paper page and confirmed by Hugging Face.

Do upvotes mean a paper is good?

They mean it got attention on the day it was posted. The feed rewards timing, a well-known author list and a striking title as much as it rewards results. Treat benchmark claims as unverified until you see them independently — an LLM leaderboard is a better comparison surface than a paper’s own table, and the API cost calculator is more useful than a paper’s efficiency claims when you are deciding what to actually deploy.

Where do the evaluation numbers on a model card come from?

From the model-index block in that repo’s README.md, which the repo owner writes by hand. They are self-reported unless a metric carries verified: true, which indicates Hugging Face ran the evaluation. Treat unverified entries the way you would treat any number in a preprint: plausible, unaudited, and worth re-running on your own data.