How to request and assess a data sample under NDA

To request a data sample under NDA, tell the seller your task, ask for a representative extract drawn the way the full set would be, agree handling rules in writing, and then test the sample for format, quality, duplication, coverage and fit before you bid. Set a deadline for yourself to decide.

Why ask for a sample at all?

A sample is the only direct evidence that a dataset matches its description. Descriptions are written by sellers, who know the data well but are not neutral, and what looks right in a summary can fail on the first file you open.

Samples also test the relationship. How quickly the seller answers, how clearly they explain gaps and whether they accept reasonable handling terms all predict how delivery of the full set will go.

There is a cost on the seller's side too. Preparing a sample takes time and carries risk, because even a small extract can contain sensitive content. A focused, respectful request gets a better sample, faster.

Timing matters as well. Ask early in the process, before you have spent engineering time on integration plans, and tell the seller when you expect to decide. A seller who knows a decision is due next week will prioritise the request over one that feels open-ended.

Why does the seller want an NDA?

Sellers want an NDA because a sample may contain confidential business information, personal data or trade secrets, and once shared it cannot be taken back. The agreement sets the ground rules for how you may use and store it.

A typical data-sample NDA covers what counts as confidential, permitted uses (usually evaluation only), who may see the sample, how it must be stored, how long the obligations last and what happens to copies afterwards. Sellers may also want a clause confirming you will not attempt to identify individuals in the sample.

Read it before you sign. Pay attention to whether it restricts your general know-how, defines evaluation narrowly enough to block your actual test, or forbids using any tools beyond your own systems. This is general information and not legal advice; have a qualified lawyer review any NDA that you are unsure of.

Terms worth negotiating

Most data-sample NDAs are reasonable, but a few terms deserve attention before you sign:

• Purpose. Make sure "evaluation" covers the tests you plan to run, including trial fine-tuning if that matters to you.
• Duration. Obligations that last forever on a trivial sample are worth questioning.
• Residual knowledge. A clause saying you may keep general learning from the sample, without copying content, protects your team's ordinary work.
• Return and deletion. Agree a clear deadline and how you will confirm deletion.
• Jurisdiction. Know which law and courts apply, particularly in a cross-border deal.

What should your request include?

A good request states who you are, what you intend to build, what the sample needs to show and how you will protect it. Seller time is limited, and specifics make it easy to say yes.

1. A short description of your task and the type of model or system involved.
2. The size and shape of sample you want, such as a few hundred records across the full date range.
3. How you want it drawn: random, stratified by category, or including known awkward cases.
4. The fields and format you expect, and whether you can accept masked or redacted fields.
5. How you will store and access it, and who on your team will see it.
6. Your timeline for a decision.
7. A request for summary statistics on the full dataset: count, date range, class balance, language mix.

On LogitAtlas, you can send this through the listing's message thread, which keeps the exchange attached to the listing. The seller can answer, counter or ask questions there, and later you can move to a bid with the history in one place.

What makes a sample representative?

A representative sample is drawn by a documented method from the whole set, not hand-picked. If the seller chooses the cleanest examples, you learn only what the best case looks like.

Ask how it was drawn. Random sampling over the full range is the baseline. If categories matter, ask for a stratified sample so rare ones appear. If you are worried about drift, ask for records from the start, middle and end of the period.

Be realistic about sizes. A sample of a few hundred records can reveal format, structure and obvious quality problems. It will not reveal rare events or subtle bias, which is why you also ask for statistics about the whole set.

How do you assess the sample?

Work through structure, quality, coverage, sensitivity and task fit in that order. Each step can end the evaluation early, so start with the cheapest tests.

Structure and format

Can you parse it with the tools you plan to use? Check encodings, schemas, delimiters, nested fields and whether the fields match the documentation. If it takes a day to load, expect cleaning costs on the full set.

Quality

Read a few dozen records by hand. Look for truncation, garbled text, empty fields, inconsistent labels and obvious errors. Measure duplication and near-duplication, because repeated records inflate apparent volume.

Coverage

Compare the sample against the seller's description. Do the date range, categories, languages and sources match? Are the cases you care about, such as unusual requests, errors or escalations, present?

Sensitivity

Search for leftover personal or confidential information: names, emails, phone numbers, account numbers, signatures, internal links. If you find any, tell the seller promptly and agree what to do with that portion. See how to verify data provenance for the wider checks.

Task fit

Run the sample through a small version of your real task. That could be a quick fine-tune, a retrieval test, a labelling pass or a comparison against your current data. Even a rough test tells you more than hours of reading.

How do you test fit in a single day?

Set up a small, repeatable test before the sample arrives, so you can run it the same day. This also stops you from adjusting the test to flatter the data.

A workable plan:

1. Prepare an evaluation set from your own data or tasks before you see the sample.
2. Write a short checklist of the structure, quality and coverage checks above.
3. When the sample arrives, run the checklist and record the results.
4. Run your small task test with and without the sample, and compare against the evaluation set.
5. List open questions for the seller, grouped as blocking and non-blocking.
6. Write one paragraph recommending whether to proceed, bid lower, ask for changes or walk away.

Hypothetical example: you are evaluating a corpus of support conversations to improve a triage model. You run the sample through your existing classifier and find that a quarter of conversations lack the final resolution message. That is a blocking question, because your model needs the outcome, so you ask whether the missing resolutions are in the full set.

Score the sample before you read the conclusion

Give each check a simple pass, concern or fail mark before you form an overall view. It stops one impressive feature, such as clean formatting, from covering a serious problem, such as missing outcomes. Share the marks with a colleague who has not seen the seller's pitch. If you both reach the same decision from the same marks, you can be more confident in it.

How should you handle the sample safely?

Treat it as sensitive regardless of how clean it looks. Store it in a restricted location, limit access to the named people, and do not paste it into external tools or services unless the NDA allows it.

Be particularly careful about hosted AI tools. Sending a sample to a third-party model for analysis can breach the NDA and may expose confidential content. Use your own environment unless the seller has agreed otherwise.

When you finish, delete the sample or return it as the NDA requires, and note the date. If you go on to buy the full dataset, the license terms replace the NDA's evaluation limits, so check what you are then permitted to do with the sample.

What if the sample reveals a problem?

Report it to the seller promptly, specifically and without blame. Describe what you found, how many records it affects and what you need to proceed. Sellers often fix a masking gap or supply a better extract quickly, and how they respond is useful evidence of how they will behave if a problem appears after payment.

If the problem is structural, such as the data missing the field your task depends on, say so plainly and end the discussion. A short, courteous no saves both sides time and keeps the door open for a better fit later.

What can a sample not tell you?

A sample cannot show rare events, long-range drift, rights issues or whether the rest of the dataset matches. Do not take a good sample as proof of a good dataset.

It also cannot show whether the seller has the right to license. For that, you need written answers and warranties, discussed in how to license enterprise data for AI training. Treat the sample as one input, and make the final decision on the combined evidence.

If the sample passes, you can move to the bid flow. On LogitAtlas, buyers bid, sellers counter or accept, and an accepted bid becomes the checkout price. Payment is held in escrow until you confirm delivery.

Frequently asked questions

Is an NDA always required for a data sample?

Not always, but it is common whenever the sample could contain confidential or personal content. Some sellers share small, clearly non-sensitive extracts without one. If a seller asks for an NDA, agree to reasonable terms and have a lawyer review anything unusual.

How big should a data sample be?

Large enough to reveal structure, quality and obvious gaps, usually a few hundred records for text data, but it depends on the data type. Pair it with summary statistics about the full set, because no small sample can show rare events.

Can I use the sample to train a model?

Usually not. Samples are normally shared for evaluation only, and the NDA or license will say so. Training on a sample without permission can breach the agreement. If you want to test training effects, ask the seller whether limited experimental use is allowed.

What if the sample is much cleaner than the full dataset?

That is a known risk. Ask how it was drawn, and request statistics and a second sample from a different part of the dataset. If the seller will not describe the method, treat that as a warning.

What if the seller refuses to provide a sample?

Ask why, and offer alternatives such as a masked extract, a schema with statistics, or a call. If the seller will provide nothing at all for a purchase of any size, think carefully about whether to proceed.

Ready to ask? Browse the datasets and open the message thread on a listing that fits your task.