# Buy source code datasets for AI training

URL: https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets

Source code datasets are licensed private codebases, ideally with their commit history, pull requests, issues and runbooks, used to train and evaluate coding models and agents. DataVendor says labs pay more for code written mostly by people and for systems languages such as Rust, Go, C and C++. MIT Technology Review reported one closed startup's codebase selling for about $25,000.

## Key takeaways

- AI teams use source code datasets for (1) pre-training and fine-tuning coding models, (2) training software engineering agents on real repository tasks and (3) building evaluation tasks from issues and the commits that fixed them.
- Usually priced per repository or codebase, as a flat license fee.
- Common delivery formats: Git repository bundles with full history, ZIP or tar archives of a fixed snapshot and JSONL of pull requests, reviews and issues.
- Check before you license source code datasets: (1) US copyright: employee code owned as work made for hire, contractor code assigned in a signed writing, (2) open-source components and their licenses identified and (3) secrets, keys and credentials removed from code and history.
- Every seller we shortlist passes 8 vetting checks before you see a sample.

## Buyer's spec

- Used for: Pre-training and fine-tuning coding models; Training software engineering agents on real repository tasks; Building evaluation tasks from issues and the commits that fixed them; Code review and debugging models trained on pull request history
- Formats: Git repository bundles with full history; ZIP or tar archives of a fixed snapshot; JSONL of pull requests, reviews and issues; Runbooks and incident notes in Markdown
- Usually priced: Per repository or codebase, as a flat license fee
- Check before you sign: US copyright: employee code owned as work made for hire, contractor code assigned in a signed writing; Open-source components and their licenses identified; Secrets, keys and credentials removed from code and history; Client-owned code excluded or cleared; Share of AI-generated code disclosed; Commit history intact and matching the manifest

## What is a source code dataset?

A source code dataset is a licensed copy of private software: the code itself plus, in the stronger packages, the history that explains it. DataVendor describes production code as private examples of how developers build, test and maintain working software, and notes that public repositories rarely preserve the full operational record.

Context adds value. DataVendor lists pull requests, issue tickets, runbooks and incident notes as code-adjacent material that explains why the code changed.

## What do labs pay most for in source code?

Labs pay most for source code written mainly by people, in systems languages, from large engineering teams. DataVendor states that labs pay more for human-written code and for Rust, Go, C and C++, and that high-quality codebases from large engineering teams sell to the most lab buyers on its platform.

| Factor | What buyers prefer |
| --- | --- |
| Authorship | Code written mostly by people |
| Language | Systems languages such as Rust, Go, C and C++ |
| Team | Codebases from large engineering teams |
| Context | Pull requests, issue tickets, runbooks and incident notes |
| History | The full operational record, which public repositories rarely keep |

## How much does a source code dataset cost?

Source code datasets range from tens of thousands of dollars to the high six figures, depending on size, history and language. The public figures are few and come mostly from the supply side.

- About $25,000 for a closed crypto startup's codebase, reported by MIT Technology Review in October 2026
- $50,000 to $500,000 for software companies with 15 to 500 employees, from Mercor's published listing for private code and its history
- $20,000 to $890,000 for design and web agencies' design systems, shipped websites and the code behind them, also from Mercor's listing

MIT Technology Review also reported that buyers' valuations rest on research theses estimating how many useful tasks a body of code might yield. When you price a codebase, ask the seller how many issues have linked fixes and tests, since that is what a task estimate counts.

## What rights problems affect source code datasets?

Source code datasets fail diligence most often over ownership and open-source licenses. Under 17 USC 201(b), code an employee writes within the scope of employment is a work made for hire owned by the employer. Code from contractors usually needs a transfer signed by the contractor under 17 USC 204(a), because the US Copyright Office's Circular 30 limits commissioned works made for hire to nine listed categories with a signed written agreement.

The seller must also identify open-source components whose licenses impose conditions. If you are buying from a UK seller, section 11 of the Copyright, Designs and Patents Act 1988 likewise leaves contractors with copyright unless they assign it in writing.

- Copyleft components such as GPL code, which carry their own license conditions
- Client-owned code written under services contracts
- Vendored third-party libraries and SDKs under commercial terms
- Secrets in commit history, which survive deletion from the current code

Every seller we shortlist passes our [vetting standard](https://sellmydatatoai.net/how-we-vet), which includes chain of title and a check that third-party content and open-source code are identified and either cleared or removed.

## How should a source code dataset be delivered?

A source code dataset should be delivered as a fixed export: a git bundle or archive of each repository with its full history, plus a separate export of pull requests, reviews and issues from the code hosting platform. Frankfurt Kurnit Klein and Selz advises against giving buyer tools direct access to company systems, and that applies to code hosts as much as to email.

Ask for a manifest per repository listing languages, lines of code, commit count, date range and number of contributors, so you can check the delivery against what you paid for.

## What should you check before buying source code?

Before buying source code, run a license scan on a sample, confirm the commit history matches the manifest and ask what share of the code was generated by AI tools. On DataVendor's account of buyer preferences, a history full of tool-generated commits is worth less.

Our guide on [how to sell source code to AI companies](https://sellmydatatoai.net/guides/sell-source-code-to-ai-companies) shows what sellers prepare, and [agent trajectories and RL environments](https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments) covers turning repositories into tasks. Software firms can read about [selling software company data](https://sellmydatatoai.net/sell-your-data/software-companies). To source vetted codebases, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer).

## Frequently asked questions

### Can AI companies buy private source code?

AI companies can buy licenses to private source code from the companies that own it. DataVendor runs a marketplace for code repositories, and Mercor publishes payouts of $50,000 to $500,000 for software companies with 15 to 500 employees. Under US copyright law the seller must own employee code as work made for hire and hold signed assignments for contractor code.

### Does AI-generated code reduce a codebase's value?

AI-generated code does reduce a codebase's value to most training buyers. DataVendor states that labs pay more for code written mostly by people, so ask the seller to estimate the share of tool-generated code and the date from which AI coding tools were used across the team.

### What about open-source code inside a private repository?

Open-source code inside a private repository must be identified before sale, because its license still applies. Copyleft licenses such as the GPL impose conditions on distribution. Our seller vetting checks that third-party content and open-source code are identified and either cleared or removed from the dataset.

### Is commit history worth including in a code dataset?

Commit history is worth including in a code dataset because it shows how and why the code changed. DataVendor notes that public repositories rarely preserve the full operational record, and that pull requests, issues, runbooks and incident notes add the context labs value.

## Sources

- [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026
- [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026
- [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026
- [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026
- [17 U.S. Code section 201: ownership of copyright](https://www.law.cornell.edu/uscode/text/17/201), Legal Information Institute, Cornell Law School, accessed November 2026
- [17 U.S. Code section 204: execution of transfers of copyright ownership](https://www.law.cornell.edu/uscode/text/17/204), Legal Information Institute, Cornell Law School, accessed November 2026
- [Works Made for Hire (Circular 30)](https://www.copyright.gov/circs/circ30.pdf), US Copyright Office, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation