# How to Publish Your Dataset Directly from a Data Sharing Platform

By the end of this guide, you will be able to choose an appropriate repository, prepare your files and metadata to a publishable standard, upload them, and walk away with a citable Digital Object Identifier (DOI, a permanent web address assigned to your dataset) that you can reference in your manuscript. You will need your raw and processed data files, a variable list, a text editor for the README and data dictionary, your ORCID identifier (a persistent author ID), your grant numbers, and a decision about the license you want to apply.

The whole process is faster than most researchers expect. The slow part is not the upload. The slow part is deciding what the dataset contains and documenting it well enough that a stranger can reuse it. That preparation is what separates a dataset that gets downloaded from one that sits unread, and this guide covers both halves.

## What Publishing a Dataset Actually Means

![Decision path asking whether data carry privacy risk, leading to open, managed or closed access models.](/article-images/illustrations/publish-dataset-from-data-sharing-platform-access-model-choice.webp)

Publishing a dataset is different from making it "available." A published dataset sits in a repository with a persistent identifier, a stated license, structured metadata, and a version history. Anyone can find it, cite it, and check whether the claims in your paper hold up against the numbers you actually collected.

The framework most repositories follow is FAIR, which stands for findable, accessible, interoperable, and reusable. The principles were laid out by Wilkinson and colleagues in 2016 and now shape how repositories design their submission systems [1]. In practice, FAIR means your dataset has a permanent identifier, a human-readable description, files in open formats, and documentation that explains what each column means.

Data sharing platforms differ in how much control you keep over access. Health data platform researchers describe three main models [2]:

1. **Open access.** Anyone can download the data after accepting a simple terms-of-use statement. This model maximizes reuse and citations.
2. **Managed access.** Requesters must submit a research proposal, sign a data use agreement, and pass a review before downloading. This suits human subjects data and other governed material.
3. **Closed or private.** Only members of a single institution or consortium can reach the data. This model exists for legal or ethical reasons but sharply limits reuse.

The model you can use depends on your data type and governance requirements, and choosing the right one is the first real decision in the process.

There is also an empirical reason to deposit data rather than promise later access. A mixed-methods study examined articles from 333 open-access journals published by BioMed Central in January 2019. Of 3,556 articles, 3,416 contained a Data Availability Statement, and the most frequent category, at 42 percent, said data were "available on reasonable request." Yet among 1,792 authors who promised to share, 93 percent either never responded to a request or declined, and only 6.8 percent actually provided the requested data [3]. A promise in a manuscript is not a publication. A deposited dataset is.

## Choosing Where to Publish: A Data Repository Comparison

You have two broad choices. A **general-purpose repository** accepts data from any field. A **domain-specific repository** is built around one community's file formats, metadata standards, and review practices.

General repositories such as Zenodo, Figshare, and ScienceDB handle almost anything and are easy to use. Their weakness is that they are not optimized for any particular data type. The developers of ShareLoc, a platform for single-molecule localization microscopy data, made this point explicitly. Repositories such as Figshare and Zenodo are generic in purpose and not optimally suited for gathering, visualizing, and exploiting their community's data in a manner consistent with FAIR principles. A single super-resolution image is computed from many thousands of raw frames that often total many gigabytes, a size that general repositories handle poorly.

Domain repositories solve that problem by knowing exactly what your data should look like. Consider the range that exists:

- **ShareLoc** hosts localization microscopy data with community-appropriate visualization and quality standards [4].
- **The Open Data Commons for Traumatic Brain Injury and Spinal Cord Injury** are secure platforms where the neurotrauma community accesses curated public data and publishes datasets with a citable DOI. They enforce minimum data standards and provide an open-source Data Quality App that automatically checks whether an uploaded dataset meets platform specifications [5].
- **SilkMeta** integrates silkworm pan-genome and multi-omics data, including 1,082 short-read genomes and 1,168 transcriptomes, into twelve functional modules with visualization and retrieval tools [6].
- **EcoBank** is an open, web-based ecological platform built on WebGIS, an application programming interface, and open-source software, with modules for [data management](/blog/guides/data-management-basics-principles-processes-and-best-practices), analysis, and service [7].
- **The EMIF Catalogue** lets biomedical data custodians publish and share different levels of information about patient cohorts while researchers search for databases that meet study requirements. It supports managed access rather than open download [8].
- **NMDMS**, a materials science repository, requires submission through a standardized data model and runs a rigorous project data review and publication subsystem before anything goes live. Since 2018 it has published more than 12 million pieces of data across 87 categories from 45 projects, with over 2.4 million downloads [9].
- **ScienceDB** is one of the largest global scientific data sharing platforms and now layers AI-based recommendation services on top of its holdings [10].
- **TOPICS-MDS** shows what a coordinated initiative looks like. Between 2010 and 2013, 41 research projects contributed data, producing records for 32,310 older persons and 3,940 informal caregivers in a single public repository [11].

The table below summarizes the main options.

| Repository type | Example | Access model | Best suited for |
|--|--|--|--|
| General purpose | Zenodo, Figshare, ScienceDB | Mostly open | Small to medium datasets in any field, quick publication, no domain standard to satisfy [4][10] |
| Domain specific, open | ShareLoc, EcoBank, SilkMeta | Open or community specific | Large files or specialized formats that need domain visualization and tooling [4][6][7] |
| Domain specific, curated | ODC-TBI and ODC-SCI | Open with automated quality checks and DOI issuance | Communities with formal data standards [5] |
| Managed access | EMIF Catalogue | Proposal review, data use agreements | Human subjects and governed clinical data [8][2] |
| Review-before-publication | NMDMS | Curated, semi-structured schemas | Heterogeneous project data that must meet a national or funder schema [9] |

A reasonable default rule: if your field has a domain repository that issues DOIs, use it. If it does not, or your dataset does not fit, a general repository is a perfectly credible choice. Reviewers care about completeness of documentation more than about which platform you picked.

## How to Publish Your Dataset from a Data Sharing Platform, Step by Step

The walkthrough below uses a running example: a CSV file of tree measurements from three field sites, the kind of table that the FOR-species20K benchmark project scaled up to more than 20,000 trees across 33 species when it published its laser-scanning dataset. The same steps apply whether your data are genomic, ecological, clinical, or behavioral.

### Step 1: Prepare and Clean the Dataset

Decide what you are releasing. Most published datasets include a processed analysis file plus whatever raw data are reasonable in size. Separate the two clearly, because a reviewer who wants to rerun your analysis needs the processed file, while a methodologist who wants to re-derive it needs the raw one.

Handle sensitive identifiers before anything leaves your lab. Anonymization is standard practice, and there is a useful pattern to copy from cybersecurity research. In one published intrusion detection dataset, network identifiers such as IP addresses and hostnames were anonymized, but the researchers supplied a list of interesting features of those entities, such as blacklist presence and geolocation, collected at the time of data capture. The anonymized release kept nearly all analytical value without exposing the original identifiers [12]. The same trick works for human subjects data. Replace names and exact addresses with coded values, then publish the codebook of derived features separately.

Check your file formats. CSV for [tabular data](/blog/research-skills/tabular-data-what-it-is-and-how-to-analyze-it), plain text for sequences, and open formats rather than proprietary ones. If your instrument writes a closed format, export a copy and describe the conversion in your README.

### Step 2: Write the Data Dictionary

A data dictionary is a table describing every variable in your dataset: its name, units, allowed values, and how it was measured. Curated repositories treat this as mandatory. An audit of the Open Data Commons platforms found that the most common reasons a dataset failed review were a missing data dictionary and mismatched column headers between the dataset and its dictionary, so much so that the platforms built an automated tool to catch these errors before submission [5].

Build the dictionary as its own small CSV file. For the tree example:

```
variable,description,unit,allowed_values
site,Field site code,none,A, B, or C
species,Scientific name of tree,none,Latin binomial
height_m,Tree height measured with clinometer,meters,positive number
dbh_cm,Diameter at breast height,centimeters,positive number
measure_date,Date of measurement,ISO 8601 date,YYYY-MM-DD
```

A useful habit is to check every variable name in the dictionary against the header row of your data file. The Data Quality App developed for the neurotrauma platforms automated exactly this comparison and gave immediate feedback on whether the dataset met specifications or needed edits, and the platforms saw a trend of fewer failed checks as the tool matured [5]. You can do a lightweight version of the same check yourself in the next step.

### Step 3: Validate Files and Generate Checksums

Before uploading, run two checks. First, confirm the data parse cleanly and the dictionary matches. Second, generate a checksum so anyone can verify later that the files were not corrupted.

```
import pandas as pd

df = pd.read_csv("tree_measurements.csv")
dict_df = pd.read_csv("data_dictionary.csv")

print(df.shape)
print(df.isna().sum())
data_cols = set(df.columns)
dict_cols = set(dict_df["variable"])
print("In data but not dictionary:", data_cols - dict_cols)
print("In dictionary but not data:", dict_cols - data_cols)
```

Expected output for a clean dataset looks like this (numbers are illustrative):

```
(420, 6)
site            0
species         0
height_m        3
dbh_cm         12
measure_date    0
In data but not dictionary: set()
In dictionary but not data: set()
```

Missing values are acceptable if the dictionary says so. Undocumented missing values are not. Fix the dictionary or the data until the two column lists are identical.

Then create checksums. A checksum is a fixed-length fingerprint of a file. If even one byte changes, the fingerprint changes.

```
$ sha256sum tree_measurements.csv data_dictionary.csv > checksums.txt
$ cat checksums.txt
b1946ac92492d2347c6235b4d2611184e4f3e5a1...  tree_measurements.csv
7d865e959b2466918c9863afca942d0fb89d7c62...  data_dictionary.csv
```

(Values above are illustrative. Your output will differ.) Include `checksums.txt` in your upload so reusers can confirm they have intact copies.

### Step 4: Assemble Metadata

Metadata is the structured description that makes your dataset findable. Most platforms present this as a web form, but it pays to draft it in advance. The fields that matter are:

- **Title.** Descriptive, not clever. "Tree measurements from three mixed forest sites in Ohio, 2025" beats "Forest data."
- **Creators.** Full names with ORCID identifiers so the dataset links to your publication record.
- **Abstract or description.** Three to six sentences covering what was measured, how, where, when, and what the known limitations are.
- **Methods.** Instrument models, sampling design, and any [preprocessing](/blog/data-analysis/what-is-data-preprocessing).
- **Keywords.** Terms a stranger in your field would search for.
- **Funding.** Grant numbers, which many funders require for compliance reporting.
- **Related identifiers.** The DOI or preprint link of the paper that uses the dataset, and vice versa.

A draft in YAML form, which nearly every platform's web form maps onto:

```yaml
title: "Tree measurements from three mixed forest sites in Ohio, 2025"
creators:
  - family_names: "Okafor"
    given_names: "Amara"
    orcid: "0000-0002-1825-0097"
description: "Field-measured height and diameter for 420 trees
  across three sites, collected May to August 2025 with a
  handheld laser clinometer and diameter tape."
keywords: ["forest ecology", "tree allometry", "field survey"]
funding: "Example Foundation grant EF-12345"
related_identifiers:
  - identifier: "10.0000/example.2026.111"
    relation: "isSupplementedBy"
```

Quality metadata is not optional decoration. The NMDMS materials repository attributes its findability and reusability directly to its standardized submission schemas and review workflow [9], and community platforms such as SilkMeta invest heavily in structured description precisely because discovery depends on it [6].

### Step 5: Choose Access Level and License

Two decisions, often confused. Access level is who can get the files. License is what they may do once they have them.

For access, revisit the three models. Fully open works when the data carry no privacy risk. Managed access is the right choice for human genomic data, identifiable clinical records, and anything your ethics approval restricts. If your funder or journal mandates open data but your data are sensitive, publish the anonymized derived dataset openly and register the identifiable raw version under managed access, describing both in your metadata.

For license, the common open choices are CC0, which waives all restrictions, and CC BY 4.0, which allows any reuse with attribution. Either is acceptable to most journals. Check your funder policy first, then pick one and state it clearly. A dataset without a license is legally ambiguous, and ambiguous datasets get reused less.

### Step 6: Upload and Publish

![Flowchart from submission through automated checks and curation review to a public record with a DOI.](/article-images/illustrations/publish-dataset-from-data-sharing-platform-upload-review-path.webp)

Uploads follow the same shape on nearly every platform. Log in, start a new deposition or submission, fill in the metadata form, drag in your files, and review the auto-generated record. Larger domain platforms may add curation steps. The Open Data Commons workflow, for example, historically involved multiple rounds of manual revision before a dataset met guidelines, and the platform's quality tool now catches most problems automatically on first pass [5]. Materials repositories with a formal review subsystem may take longer because a human reviews the submission against the schema [9].

Practical tips drawn from how successful datasets get published:

- Compress large folders into a single archive with a `README.txt` at the top level so the contents are understandable even before extraction.
- For data types with huge raw files, such as gigabyte-scale microscopy frames, deposit the compact derived product (for localization microscopy, the molecular coordinate files rather than the raw frames) and describe how the raw data can be regenerated or requested [4].
- Set an embargo date if your paper is still under review and the journal requires it. Most general repositories let you reserve the DOI now and open access later.
- Preview the public page before clicking publish. After publication, most repositories allow new versions but do not allow silent edits, which is the point of a persistent record.

Expected result: a public landing page showing your title, authors, abstract, license, file list, and a DOI in a format such as `10.5281/zenodo.XXXXXXX` for Zenodo-style depositions. Click it in a private browser window to confirm it resolves without a login.

### Step 7: Link, Cite, and Version

Publication is not the finish line. Three follow-ups make the dataset count:

1. **Cite the dataset in your paper.** Add the DOI to the Data Availability Statement and reference list using the citation format the repository provides.
2. **Link back from the dataset to the paper.** Fill in the related-identifier field so the two records point at each other. This improves discoverability on both sides.
3. **Version deliberately.** If you correct an error or add sites, publish a new version rather than deleting the old one. The old DOI keeps resolving to the superseded version, and the new DOI points to the current one. Benchmarks such as FOR-species20K are designed explicitly so that later work can build on a stable, versioned release.

## Writing the Companion Data Paper

Many researchers pair the repository deposit with a short "data descriptor" article, often in a data journal, that describes the dataset's creation in detail. The genre has a recognizable shape across fields. A Data in Brief article on intrusion detection alerts described the collection window, the contributing sources, the storage format, the categorization taxonomy, and the intended reuse cases such as correlation analysis and attack reconstruction [12]. Another Data in Brief paper on peer-to-peer lending data described over two million observations, provided variable descriptions, descriptive statistics, and accompanying analysis code, and spelled out the longitudinal questions the panel dataset could support [13]. A mobility data paper described the scraping interval, the clustering method used to identify idling positions, how trips were inferred, and the validation against official records [14].

The template generalizes. A data paper tells a reader what is in the files, how the data were produced, what quality checks were run, and what the data can and cannot be used for. It does not present new scientific hypotheses or results. Data papers carry their own DOIs and citations, which means the dataset earns scholarly credit twice, once as a repository record and once as a paper.

Before writing one, check whether your target data journal requires the dataset to already be published in a repository. Most do. That ordering, repository first, paper second, is also the order that guarantees your manuscript can cite a working DOI.

## Common Mistakes and Limitations

These are the errors that most often force a revision, or worse, leave a dataset unusable.

**Relying on "available on request."** The compliance study cited above found that a Data Availability Statement promising data on request performed no better than providing no statement at all, with 93 percent of promising authors failing to share when asked [3]. Deposit the files.

**Skipping the data dictionary.** Missing dictionaries and mismatched headers were the dominant failure modes in a large audit of curated repository datasets [5]. If your dataset has columns, it needs a dictionary.

**Uploading undocumented raw data as the only artifact.** A processed file with the dictionary is more useful than hundreds of gigabytes of raw instrument output with nothing explaining it. For fields with very large raw files, publish the compact derived product and describe the raw pipeline.

**Forgetting anonymization.** Identifiers in raw operational data, whether IP addresses or patient records, must be removed or coded before open release. The pattern of publishing anonymized identifiers alongside a preserved feature list is a proven middle path [12].

**Choosing a repository by habit instead of fit.** A generic repository is fine for modest datasets, but specialized communities build repositories because generic ones cannot visualize or standardize their data types. Conversely, a managed-access biomedical platform is the wrong place for an open dataset that needs no gatekeeping [8].

**Mixing up access restrictions with licensing.** Restricting download does not remove the need for a license on whatever is accessible. State both.

**No version plan.** Silently replacing files after publication breaks every citation that pointed at the original record. Publish a new version and keep the old one.

**Treating submission as the moment quality gets decided.** Curated platforms will bounce a dataset back, sometimes repeatedly, when headers, schemas, or dictionaries fail checks [5][9]. Running your own validation before upload, as in Step 3, prevents most round trips.

One general limitation applies to any dataset you publish. Sharing enables reuse, but reuse still requires domain judgment. People integrating your field data, clinical measurements, or model outputs need their own expertise to interpret them correctly, and your job is to give them the documentation to do so.

## Frequently Asked Questions

### What does it mean to publish a dataset from a data sharing platform?

It means depositing your data files with structured metadata in a repository so the deposit receives a persistent identifier, usually a DOI, and becomes permanently findable, accessible, and citable under a stated license.

### Which repository should I choose for my dataset?

Choose your field's domain repository if one exists and issues DOIs, because it will handle your file formats and standards properly. Otherwise, a general repository such as Zenodo, Figshare, or ScienceDB is a credible and widely accepted choice.

### What files should I include when I publish a dataset?

Include the processed analysis file, a data dictionary describing every variable, a README explaining contents and provenance, a checksum file, and raw data when size and sensitivity allow it.

### What metadata does a repository require?

At minimum you will provide a descriptive title, creator names with ORCID identifiers, an abstract, methods, keywords, funding information, license, and any related publication identifiers.

### Do I need to write a separate data paper?

No, it is optional, but a short data descriptor in a data journal documents provenance, gives the dataset a second citable record, and typically requires the repository deposit to exist first.

### Can I publish sensitive or human subjects data this way?

Yes, but use a managed access model, where requesters pass a review and sign a data use agreement before download, or publish an anonymized version openly and keep the identifiable source restricted.

### What is a DOI and why does my dataset need one?

A DOI is a permanent identifier that resolves to your dataset record no matter where the files are hosted, which lets others cite your data the way they cite a paper and lets you list it on your CV.

### How do I update a dataset after it has been published?

Publish a new version through the repository rather than editing files in place, so the original DOI keeps resolving to the initial record and the new version carries its own DOI linked to the first.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What does it mean to publish a dataset from a data sharing platform?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It means depositing your data files with structured metadata in a repository so the deposit receives a persistent identifier, usually a DOI, and becomes permanently findable, accessible, and citable under a stated license."
      }
    },
    {
      "@type": "Question",
      "name": "Which repository should I choose for my dataset?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Choose your field's domain repository if one exists and issues DOIs, because it will handle your file formats and standards properly. Otherwise, a general repository such as Zenodo, Figshare, or ScienceDB is a credible and widely accepted choice."
      }
    },
    {
      "@type": "Question",
      "name": "What files should I include when I publish a dataset?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Include the processed analysis file, a data dictionary describing every variable, a README explaining contents and provenance, a checksum file, and raw data when size and sensitivity allow it."
      }
    },
    {
      "@type": "Question",
      "name": "What metadata does a repository require?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "At minimum you will provide a descriptive title, creator names with ORCID identifiers, an abstract, methods, keywords, funding information, license, and any related publication identifiers."
      }
    },
    {
      "@type": "Question",
      "name": "Do I need to write a separate data paper?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "No, it is optional, but a short data descriptor in a data journal documents provenance, gives the dataset a second citable record, and typically requires the repository deposit to exist first."
      }
    },
    {
      "@type": "Question",
      "name": "Can I publish sensitive or human subjects data this way?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Yes, but use a managed access model, where requesters pass a review and sign a data use agreement before download, or publish an anonymized version openly and keep the identifiable source restricted."
      }
    },
    {
      "@type": "Question",
      "name": "What is a DOI and why does my dataset need one?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "A DOI is a permanent identifier that resolves to your dataset record no matter where the files are hosted, which lets others cite your data the way they cite a paper and lets you list it on your CV."
      }
    },
    {
      "@type": "Question",
      "name": "How do I update a dataset after it has been published?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Publish a new version through the repository rather than editing files in place, so the original DOI keeps resolving to the initial record and the new version carries its own DOI linked to the first."
      }
    }
  ]
}
</script>

## Sources

1. [Toward the Open Science model: publish your raw diffraction data.](https://doi.org/10.1063/4.0000368)
2. [Health Data Sharing Platforms: Serving Researchers through Provision of Access to High-Quality Data for Reuse](https://pubmed.ncbi.nlm.nih.gov/38487482/)
3. [Many researchers were not compliant with their published data sharing statement: mixed-methods study.](https://pubmed.ncbi.nlm.nih.gov/35654271/)
4. [ShareLoc - an open platform for sharing localization microscopy data](https://pubmed.ncbi.nlm.nih.gov/36271059/)
5. [Improving Data Quality in Neurotrauma Data Sharing: The Open Data Commons Minimal Data Standards and a Data Tool for Compliance.](https://pubmed.ncbi.nlm.nih.gov/42157403/)
6. [SilkMeta: a comprehensive platform for sharing and exploiting pan-genomic and multi-omic silkworm data.](https://pubmed.ncbi.nlm.nih.gov/37941143/)
7. [EcoBank: A flexible database platform for sharing ecological data](https://pubmed.ncbi.nlm.nih.gov/33654452/)
8. [EMIF Catalogue: A collaborative platform for sharing and reusing biomedical data](https://pubmed.ncbi.nlm.nih.gov/31029262/)
9. [A repository for the publication and sharing of heterogeneous materials data](https://pubmed.ncbi.nlm.nih.gov/36575234/)
10. [ScienceDB AI: An LLM-Driven Agentic Recommender System for Large-Scale Scientific Data Sharing Services](https://doi.org/10.1145/3770855.3818479)
11. [The Development of the Older Persons and Informal Caregivers Survey Minimum DataSet (TOPICS-MDS): A Large-Scale Data Sharing Initiative](https://pubmed.ncbi.nlm.nih.gov/24324716/)
12. [Dataset of intrusion detection alerts from a sharing platform](https://pubmed.ncbi.nlm.nih.gov/33299907/)
13. [FinTech and macroeconomics: Dataset from the US peer-to-peer lending platform](https://pubmed.ncbi.nlm.nih.gov/34926742/)
14. [High-resolution dataset of car, bike, and e-scooter sharing in Munich, Germany (2023-2025): Vehicle idling positions and derived trips](https://pubmed.ncbi.nlm.nih.gov/41140854/)