Research Data: Opensource Datasets from Global Software Companies

At Global Software Companies (GSC), we publish the data behind our research so you can check our figures or run your own analysis. Every dataset here is free to download and reuse, commercially too, under CC BY 4.0. Each comes with a field guide that explains every column and a script that recomputes the figures in our articles. Get them from whichever platform you already use: Hugging Face, GitHub, Kaggle or Zenodo, which keeps the permanent archived copy behind each DOI. Every copy has the same data files. We add new datasets here as we publish them.

Software Job Listings Dataset

US entry-labeled software job listings shows what 2,517 US software jobs labeled “Entry level” on LinkedIn actually asked for. They were posted between January and August 2026, most of them in July and August.

What’s in it

• jobs.csv: one row per listing, with 75 columns covering the experience required, degrees and the alternatives employers accepted, citizenship, clearance and sponsorship requirements, remote or on-site work, the projects applicants were asked to show, training on offer, how candidates are assessed, and AI use. The file has 2,518 rows; use the 2,517 marked inclusion_status = retained.

• technologies.csv: 32,271 mentions of languages, frameworks, cloud platforms, databases and AI tools, each tagged as required, preferred, optional or part of the job’s duties. Join it to jobs.csv on job_id.

• Documentation: FIELD-GUIDE.md, REPRODUCE.md, reproduce.py, LICENSE.md, manifest.json.

Which listings are included. Listings labeled “Entry level” with “software” plus developer, engineer or programmer in the title, where at least one way in asked for no more than 2 years of professional experience, or didn’t give a number. We left out internships, training placements and AI-evaluation roles, and counted repeat postings (same text, advertiser, title and location) once.

Used in: Do you need a degree for an entry-level software job?

Get the data: Hugging Face · GitHub · Kaggle · Zenodo. Whichever copy you use, please cite the DOI below.

Entry-Level Screening File

US entry-labeled software listings: screening file covers all 5,625 US software listings labeled “Entry level” that we screened, before we narrowed them down for the jobs dataset, and records how much experience each one asked for. Use it to check the figures in our article on how much experience “entry-level” software jobs ask for.

What’s in it

• screening.csv: one row per listing, showing the experience required, any sign in the listing itself that the role is junior, the type of role, whether it was AI-training work, the state, and how closely we reviewed the experience value. Rows that also appear in the jobs dataset link to it through release_job_id. For listings outside the jobs dataset, we don’t release the text, title, advertiser, location, date or pay.

• aggregates.csv: every figure published in the article, with its numerator, denominator and population. That includes figures we can’t release listing by listing, such as pay medians.

• Documentation: FIELD-GUIDE.md, REPRODUCE.md, reproduce.py, LICENSE.md, manifest.json.

About the figures in our article

• The 18.0% requiring 2+ years holds up. Across the ways we checked it, the share stayed between 16% and 20%: 19.8% counting identical descriptions once (756 of 3,811), 19.5% without the advertiser that posted the most listings, 18.6% for plain software-engineering roles, 16.7% for the US software roles within our scope, and 16.5% without AI-training listings (854 of 5,187).

• What counted as experience. We only counted a number of years when the listing framed it as professional, industry or work experience, or as software development experience in general. If you also count years tied to one language, tool or area of software work, the share rises to 20.7%.

• “Twice as many.” For every listing that explicitly accepted no experience, 2.06 required 2+ years (1,010 against 490). Without AI-training listings the ratio is 1.74; counting identical descriptions once, it’s 2.42; and for plain software-engineering roles, it’s 2.69.

• Pay. Where a listing gave a yearly pay range, we used the midpoint; where it gave a single yearly figure, we used that. We didn’t count pay mentioned only in the description. The “Entry level” and “Mid-Senior level” medians come from the 1,055 and 7,003 listings with yearly pay in our wider collection of 96,309.

Used in: We analyzed 5,625 “entry-level” software jobs

Get the data: Hugging Face · GitHub · Kaggle · Zenodo. Whichever copy you use, please cite the DOI below.

CS Degree AI Requirements Dataset

Selected CS degree AI and ML requirements records whether computer science degrees require a course in artificial intelligence or machine learning, based on each university’s published degree rules. It covers 25 US tracks at 21 universities and 21 European programs, each at two points in time: 2021–22 and 2026–27.

What’s in it

• universities.csv: 92 rows, one per program per year, with 30 columns: whether a course named for AI or ML was unavoidable, which AI or ML courses could count toward the degree, optional routes, credits, and the catalog or handbook behind each judgment, with links and page references.

• Documentation: FIELD-GUIDE.md, REPRODUCE.md, reproduce.py, LICENSE.md, manifest.json.

Read this first. We count a course only if its title names AI or machine learning, so AI taught inside courses with other titles isn’t captured. We chose the programs deliberately, partly because their documents were available, so treat them as case studies rather than a sample of all degrees. Compare the US and European groups side by side; don’t rank one against the other.

Get the data: Hugging Face · GitHub · Kaggle · Zenodo. Whichever copy you use, please cite the DOI below.

Methods

Job listings. We started with 96,309 US LinkedIn listings posted between January and August 2026 with software, developer or programmer in the title. They aren’t spread evenly: about 4 in 5 were posted in July and August, and January has very few, so the data can’t show trends over time. From these, we took the 5,625 labeled “Entry level” whose titles contained “software” plus developer, engineer or programmer, leaving out titles with senior, manager, director, instructor, tutor, sales or recruiter.

Degree programs. We read each program’s published degree rules for 2021–22 and 2026–27, using archived catalogs for the earlier year. On October 12, 2026, we checked every US 2026–27 row against the university’s live catalog. Where we couldn’t settle a program’s rules for a given year, we mark it as unresolved; we never count it as “not required.”

Classification. We classified listings and degree rules with AI assistance, following written rules that are published in each field guide. Uncertain and high-impact cases were checked again against the saved source, and any disagreements were settled. A GSC researcher checked about 1 in 5 of the AI classifications by hand and reviewed the edge cases those checks turned up. Some rows may still be wrong; when we fix them, we release a new version.

What these data are not. They’re selected listings and programs, not a representative sample of US jobs or degrees. A listing isn’t a vacancy or a hire: one employer can post the same job many times, so we report counts both ways. A requirement in a listing tells you what was asked, not how applicants were screened. Degree rules tell you what was published, not what students learned.

What we left out. We don’t release the original listing text or contact details. University documents keep their own rights, so we link to them rather than include them.

Reuse and Credit

You’re free to copy, share, adapt and build on these datasets for any purpose, including commercial use and machine learning, under CC BY 4.0. Just credit Global Software Companies, link to this page if you can, keep the license notice, and say if you changed the data. Please cite the version you used; you’ll find a suggested citation, with its DOI, under each dataset.