Insights · 6 min read
Raw Data vs Prepared Data Value: What Preparation Really Adds
Buyers do not pay for files. They pay for what they can safely train on.
Two companies can own nearly identical archives and get very different offers. The difference is rarely the material. It is the work between the archive and a buyer’s training run. This article explains what that work consists of, who normally pays for it, and why that decides how much of the value you keep.
What “raw” means
Raw data is what you can export today: a repository with its history, a chat workspace dump, a ticket system export, a folder of documents. It is real and it is yours to sell, if you have the rights. But it comes with personal information, secrets, noise and no sign of what it teaches.
What “prepared” means
Prepared data has been cleaned of secrets, had people replaced with stable pseudonyms, been linked so a change in code connects to the discussion and ticket behind it, and in the strongest case been turned into tasks with automatic checks. Each step removes work or risk from the buyer.
| Step | What it does | Evidence a buyer cares |
|---|---|---|
| Scrubbing | Removes personal data, secrets and HR records | Suppliers and buyers both scrub, and HR records are left out (MIT Technology Review) |
| Certification | Third party confirms what was removed | In one bankruptcy sale a firm must certify to the court before each transfer, including how many people’s details were removed |
| Provenance | Records source, dates, processing, removed subsets | Now treated as a standard request in AI data licenses (practitioner memo) |
| Linking | Connects code, chat, tickets and reviews into episodes | Buyers value data by the tasks it yields (MIT Technology Review) |
| Task building and checks | Turns episodes into graded, repeatable tasks | Quality criteria buyers name: soundness against reward hacking, a workable pass rate (Epoch AI) |
The best public example of the work involved
The SWE-Bench Pro paper built a commercial test set from 18 proprietary startup repositories, yielding 276 problems. According to the paper, each problem got human-written problem statements and requirements, environments built by professional engineers, and a review of each fail-to-pass test for relevance and breadth, with irrelevant or overly broad tests dropped. The authors say they contacted the company and purchased its engineering repos, though the abstract describes partnership agreements with early-stage startups, so the exact arrangement is not stated consistently.
The point is not the price. It is that a buyer who starts from raw repositories faces a long list of expensive steps. If you do those steps, or have them done before you sell, you are selling the finished product. If you do not, you are selling the ingredients, and the buyer prices in their cost and risk.
Why untested material is risky for a buyer
Epoch AI’s interviewees wanted a minimum pass rate of around 2 to 3 percent, or at least one success in 64 or 128 attempts, because a task nobody can solve may be infeasible. Tasks near a 70 percent pass rate may be retired. One neolab researcher said soundness matters most: high reward must mean the task was actually solved, not hacked. A raw archive carries none of this assurance.
Compute makes the point sharper. Epoch cites an estimate of about $2,400 of compute spent per task during reinforcement learning training. A buyer who spends that on a flawed task has wasted far more than the task cost to buy.
A before-and-after picture
Imagine one archive in two states. Before: a repository export, a chat dump with real names and a few pasted passwords, a ticket export, no documentation of rights. A buyer must find the secrets, decide what is lawful, link the pieces and invent tasks, and they will discount for the risk of missing something. After: secrets removed, people replaced by stable pseudonyms, chat and tickets tied to the commits they explain, a short provenance note, and a set of candidate tasks with checks that a reviewer has run. The material is the same. The buyer’s job is a fraction of the size, which is the value you captured by doing the work.
What preparation cannot fix
Preparation does not create history that was never recorded, and it does not create rights you do not have. It also has limits on privacy. As the founder of a digital policy center told Gizmodo about sales of workplace archives, “It’s not generic data. It’s identifiable people.” Linking records so they tell a story can make that sharper, not softer, which is why removing personal data and reviewing rights are separate jobs and both are needed.
Who should do the preparation
- You, if you have the engineers and the time. Be realistic: removing personal data and building checks is specialist work, and a mistake can expose you.
- A buyer, who will often do it but will price the cost, the risk and the delay into the offer.
- A specialist preparer such as Bearish AI, who cleans, pseudonymizes, links and grades data so it sells for more. Whoever does it, a separate review of rights, confidentiality and licensing is still needed. Removing personal data alone does not make data compliant.
How much is the gap?
We know of no published multiplier, and we would not trust one. The honest framing is that preparation moves your archive from a flat package priced on the buyer’s guess toward units a buyer can check. Small or messy archives gain the least, because there is little to link or grade.
The free estimate at /estimate shows a raw range and a prepared range side by side for your inputs, both deliberately conservative. For where those ranges sit against reported deals, see /insights/how-much-do-ai-companies-pay-for-training-data. Have a lawyer review rights before any sale.
Questions we get
- Is raw company data valuable to AI companies?
- It can be, if you hold the rights, but buyers must clean and test it first and will price that work in. Reported archive sales are modest, from about $10,000 to $100,000 for shut-down companies.
- What does preparing data for AI training involve?
- Removing secrets and personal data, replacing names with stable pseudonyms, linking records like code, chat and tickets, documenting provenance, and where possible turning episodes into tasks with automatic checks.
- Does removing personal information make data safe to sell?
- No. It is necessary but not sufficient. Ownership, customer confidentiality and licensing terms need a separate review by a lawyer.
- Should I prepare my data before approaching buyers?
- Often yes, at least a first pass: an inventory, a rights check and a scrub. Buyers expect warranties and a provenance record, and a prepared archive shortens their work.
What is your data worth?
Sources
- MIT Technology Review, “AI labs are buying data from dead startups to train their models” (October 13, 2026)
- The Next Web, “Spirit’s privacy ombudsman backs Google’s $10m data deal” (October 9, 2026)
- Terms.Law Insights, memo on a usable AI training data licensing agreement
- Epoch AI, “An FAQ on reinforcement learning environments” (January 2026)
- Deng et al., “SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?” (2025)
- SWE-Bench Pro, full text (v2), sections on commercial repositories and task verification