Insights · 6 min read
How Data Preparation Increases Data Value for AI Buyers
Buyers do not price your storage bucket. They price what they can train on, test with and defend in diligence.
Owners often ask what their data is worth as it sits. The more useful question is what it is worth once a buyer can use it. The gap between those two numbers is the value of preparation, and the evidence for why it exists is public. This article walks through it.
Buyers price the yield, not the size
MIT Technology Review reports that one buyer of company archives values them through what it calls research theses: estimates of how many useful training tasks a body of data will produce, for example per 100,000 lines of code. It also reports that buyers quote very different numbers because those estimates are uncertain. Anything that makes the yield clearer or larger raises the price you can defend.
A raw archive loses most of its material on the way to tasks
The SWE-bench team built a well-known coding benchmark from about 90,000 pull requests across 12 repositories. They kept merged pull requests that resolved an issue and changed tests, then ran the tests before and after the fix. Instances with no test that moved from fail to pass, or with installation or runtime errors, were dropped. The final set had 2,294 instances. That is roughly 2.5 percent of what they started with, drawn from 12 popular open-source Python projects.
Your own history will not match those numbers. The shape is what matters: a small share of raw records is usable as a graded task, and finding that share is work somebody must do. Interviewees in Epoch AI’s 2026 study of reinforcement learning environments mostly cited about $200 to $2,000 per task, with roughly $20,000 per task described as rare, mainly for complex software engineering.
What each preparation step changes
| Step | What it changes for the buyer | Public evidence |
|---|---|---|
| Scrubbing and pseudonymization | Makes the data possible to use at all, and shows the seller handled people responsibly | Sellers and buyers both scrub, and HR records are left out (MIT Technology Review) |
| Deduplication | Less wasted training, less memorization, cleaner tests | About ten times less memorized text emitted, and more than 4 percent train-test overlap found in standard validation sets (Lee et al.) |
| Curation for quality | Smaller, higher-quality sets can punch above their size | In the LIMA study, a model fine-tuned on 1,000 carefully curated examples produced answers human raters judged equal to or better than GPT-4’s in 43 percent of cases; a code model trained on 6 billion tokens of “textbook quality” web data plus 1 billion synthetic tokens reached 50.6 percent on HumanEval |
| Linking records | Shows why decisions were made and how bugs were found and fixed, and lets tasks carry their context | Not independently measured in public studies; buyers describe variety and depth as what raises value (MIT Technology Review) |
| Grading and verification | Turns history into tasks with a pass or fail check | Reward hacking resistance is the top quality criterion named in interviews; difficulty needs calibration (Epoch AI) |
| Documentation | Lets a buyer’s legal and research teams approve faster | EU rules now require model providers to publish summaries of training content (EU AI Act Article 53) |
Why grading is where the biggest gap opens
Epoch AI’s interviewees say tasks should be hard but solvable. They mention a minimum pass rate of about 2 to 3 percent, or at least one success in 64 or 128 attempts. Tasks that nothing solves may be infeasible, tasks solved every time give the model nothing to learn from, and tasks near a 70 percent pass rate may be retired. Someone has to run the tasks to find out. A buyer who must do that work themselves will price it in. A seller who has already done it, with evidence, takes that discount off the table.
Why linked context is worth more than the parts
MIT Technology Review reports that buyers see value in a range of situations, including disagreements and tricky judgment calls, and that real archives avoid errors common in synthetic data, such as references to dates that do not exist. It adds that the size of the performance gain is unclear and that no public comparison tests exist. The logic is plain: a chat thread that explains why a fix was chosen, linked to the ticket and the commit, is a different object from three separate files. Linking is preparation, and it has to survive pseudonymization, with the same stand-in for the same person in every source.
Three archives of the same size
Picture three hypothetical archives with the same volume. The first is a raw export with unknown rights, live credentials and real names. The second has been scrubbed and pseudonymized but sits in unconnected folders. The third has been scrubbed, deduplicated, linked into episodes, documented and partly converted into verified tasks with pass or fail checks. A careful buyer discounts the first for risk and cleanup, trusts the second a little more, and can start training or evaluating on the third in days. Same bytes, different usable yield.
Quality beats volume only up to a point
Preparation does not rescue everything. If you cannot show clear rights, no amount of cleaning makes the data sellable. If the archive is thin, there is less for preparation to work with. Reported sales of shut-down startups’ archives have been put at $10,000 to $100,000 per company. Those figures come from press accounts and measure different things from per-task prices, so they cannot be compared directly. They do show that the unit you sell, an archive or a set of verified tasks, shapes the price.
What we can and cannot say
We know of no public study that isolates how much preparation adds to a deal price, and we will not invent one. The honest version is that preparation removes discounts: for rights risk, for cleanup cost, for uncertainty about yield. It also opens routes that raw data cannot take, like selling graded tasks. The free estimator at /estimate shows two conservative ranges side by side, one for your data as it is and one for the data after preparation, so you can see the size of that gap for your own situation before spending anything.
For the order of work, read /insights/prepare-company-data-for-ai-labs. For turning bug history into tasks, read /insights/bug-history-to-fail-to-pass-coding-tasks.
Questions we get
- Does cleaning data really increase its price?
- Cleaning removes the discounts a buyer would otherwise apply for cleanup cost, privacy risk and uncertainty. Public research also shows that deduplication and curation change what models learn. No public study isolates a precise price effect, so treat any exact multiplier from anyone with caution.
- What is data preparation for AI training?
- The work of turning raw records into something a lab can use: checking rights, removing secrets, pseudonymizing people, deduplicating, linking related records, documenting the result, and where possible turning history into graded tasks.
- Is raw company data worth anything to AI labs?
- It can be, particularly when it is rare or comes with context, but raw data carries the most uncertainty about rights, privacy and usable yield. Buyers described by MIT Technology Review scrub the data themselves, and that cost shapes what they offer.
- Which preparation step adds the most value?
- It depends on the data, but turning real history into tasks with automatic pass or fail checks is the step that opens the widest gap, because it changes what is being sold from an archive to graded training material.
What is your data worth?
Sources
- MIT Technology Review, “AI labs are buying data from dead startups to train models” (October 2026)
- Jimenez et al., “SWE-bench: Can Language Models Resolve Real-World GitHub Issues?” (2023)
- Epoch AI, “An FAQ on reinforcement learning environments” (January 2026)
- Lee et al., “Deduplicating Training Data Makes Language Models Better” (2021)
- Zhou et al., “LIMA: Less Is More for Alignment” (2023)
- Gunasekar et al., “Textbooks Are All You Need” (2023)
- Gizmodo, “Failed companies are selling old Slack chats and email archives to train AI” (April 2026)
- EU AI Act, Article 53 (obligations for providers of general-purpose AI models), as published on artificialintelligenceact.eu