AI uses for records
Clean data sampling bias: a cross-test of what your cleaning removed
By SourceX Editorial · Updated
9 sections
Short answer
Clean data sampling bias appears when records are kept because they are complete, well documented or neatly resolved, since those traits track easier cases, familiar customers and good outcomes. Test it reproducibly: draw a population sample beside the clean subset, compare the two by stratum on completeness, difficulty, customer type and outcome, then train on each and score both on a representative holdout. Keep cleaning that fixes errors; reverse cleaning that removes ambiguity.
Evaluation split
Clean data sampling bias
Key takeaways
- Cleaning rules that drop whole records are sampling rules: complete fields, resolved status and confident labels all track easier cases and better outcomes.
- Measure the gap by stratum first; a share ratio near zero for reopened or refunded cases cannot be fixed by reweighting.
- The decisive test trains size-matched clean and population arms and scores both on a representative holdout as well as a curated test.
- Public benchmarks show cleaning cuts both ways: corrected labels make scores more trustworthy, while removing ambiguous or underspecified cases narrows what scores mean.
- Keep error fixes, turn messiness filters into flags, and keep one holdout that no quality rule has touched.
Why do the cleanest records misrepresent the work?
Because the traits that make a record clean are rarely independent of the work it describes. A ticket with every field filled, one owner and a tidy resolution is usually an easy ticket, from a customer type the team handles often, with a good outcome. Keep only records like that and the dataset over-represents routine success and under-represents the hard, messy and failed cases a deployed model will meet.
This is clean data sampling bias: a selection effect created by quality rules rather than by deletion, routing or a supplier's choice. NIST Special Publication 1270, on identifying and managing bias in AI, ties statistical bias to systematic error from samples that do not represent the population and notes that it can occur without any prejudice or intent. A cleaning script written in good faith fits that description exactly.
Public text data already shows the pattern. Gururangan and colleagues (EMNLP 2022) tested the quality filter used to select GPT-3 training text on U.S. high school newspaper articles and found that articles from larger schools in wealthier, more educated and urban ZIP codes were more likely to pass as high quality, while the filter's notion of quality did not line up with factuality or literary acclaim. Business records have their own version of that tilt, and it can be measured.
This page gives a reproducible test: compare the clean subset with a population sample, train on each, and score both on a representative holdout. It complements audits of deleted records, escalation queues and hand-picked supplier samples, which deal with other routes to the same problem.
Which cleaning rules quietly act as sampling rules?
Any rule that keeps or drops whole records on a trait correlated with difficulty, customer or outcome is a sampling rule, whatever the pipeline calls it. Fixing a malformed date in place changes nothing about coverage; dropping every record with a missing field can change a great deal.
The rules below are common in exports from help desks, CRMs, field service tools and ERP systems. Each is reasonable for a demo and risky for a dataset meant to stand in for real operations. The default action column is where to start before the cross-test says otherwise.
| Cleaning rule | What it keeps | What it tends to drop | Likely skew | Default action |
|---|---|---|---|---|
| Complete required fields only | Records where every mandatory and custom field is filled | Rushed, after-hours and escalated cases; records from before a field existed | Easier cases and newer periods | Flag missing fields; do not drop |
| Resolved or closed status only | Cases with a final resolution code | Open, abandoned, cancelled and reopened cases | Better outcomes than reality | Keep all statuses; label case maturity |
| Single owner or short thread | Cases one person handled in a few messages | Reassigned, multi-team and long-running cases | Lower difficulty | Keep; stratify by handoff count |
| Confident or agreed labels only | Records where labelers or rules agree | Genuine edge cases along with mislabeled ones | Fewer boundary decisions | Review; drop only confirmed errors, mark the rest ambiguous |
| Template-conforming text | Notes that parse into a standard structure | Free-text notes, workarounds and unusual requests | Uniform language, fewer exceptions | Keep raw text beside parsed fields |
| Outlier removal on amounts or durations | Values inside a set range | Large claims, long jobs, disputes | A truncated tail where costs concentrate | Cap for training only if the tail stays in testing |
| Well-specified tasks only (evaluation sets) | Requests that state every needed detail | Vague or incomplete requests that need clarification | Overstated performance on real requests | Keep an underspecified slice in the test set |
What public benchmarks show about cleaning away ambiguity
Cleaning an evaluation set makes its scores more trustworthy and, at the same time, narrows what they measure. Public benchmarks show both effects, which is why the useful question is never clean or not clean, but which cleaning, measured against what.
SWE-bench Verified is the clearest case. Released on August 13, 2024 by the SWE-bench team with OpenAI, it came from a screening in which developers experienced in Python reviewed 1,699 random samples of the original test set for underspecified issue descriptions and for unit tests that reject valid solutions; the result was a 500-task subset. The screening removed unfair tasks, and it also removed the kind of vague issue report engineers receive every day, so a high score describes performance on well-specified issues.
OpenAI's GDPval (2025), with 1,320 expert-written tasks across 44 occupations, says its tasks are one-shot and that future versions will move toward more interactive, context-rich work. The accompanying paper reports a variant with shorter, under-specified prompts on which models performed worse, mainly because they struggled to work out the missing context. A test set cleaned until every request is complete flatters models on exactly that skill.
Labels tell the same two-sided story. Northcutt, Athalye and Mueller (NeurIPS 2021) estimated label errors averaging at least 3.3 percent across 10 popular test sets and showed that such errors can change which model ranks first, so fixing genuine errors matters. Swayamdipta and colleagues' Dataset Cartography (2020) found that hard-to-learn examples are often mislabeled, while training on ambiguous examples, the ones a model keeps changing its mind about, improved out-of-distribution performance with little or no in-distribution change. A confidence filter that cannot tell those two groups apart throws away the second with the first.
Anthropic's guidance on agent evaluations adds the grader's side: checks should be inferable from the task description, so an agent is not failed for something the prompt never said, and the post describes a model that failed a task as written while finding a better solution for the user. Both points argue for keeping ambiguous cases with honest labels, not for deleting them.
The cleanliness-versus-representativeness study design
The study answers one question: on this task, does training on the clean subset beat training on a population sample of the same size, when both are scored on records that look like real operations? It uses two training arms, two test sets and a stratum key, and it reports results for each stratum as well as in total.
Fix every choice below before the first model is trained, and file it with the date and the names of the people who signed off. The design is an editorial method, version 1.0 dated 2026-10-26. SourceX has not run it on any dataset and publishes no results here; the point is that anyone, buyer or supplier, can run it and get the same answer from the same frame and seed.
| Component | Specification | Why it matters | Output to file |
|---|---|---|---|
| Question and population | The task and the operating population: systems, period, channels and customer types the model will serve | Without a population, representative has no meaning | One-paragraph population statement |
| Sampling frame | Every eligible record before cleaning, with identifiers, stratum labels and dates | The only baseline the cleaning cannot touch | Frame file, row count and hash |
| Arm C: clean-selected | Records passing the current cleaning rules, drawn to a fixed size N | Shows what the cleaning policy produces | Rules applied, with records removed by each |
| Arm P: population-sampled | Seeded random or stratified draw of size N from the frame; errors fixed in place, nothing dropped for messiness | Separates selection from sample size | Seed, draw script and stratum shares |
| Test set R: representative holdout | Seeded draw from a later period of the frame, never selected by quality, labeled by reviewers blind to arm | Measures the work the model will actually see | Holdout IDs, labeling guide, agreement rate |
| Test set K: curated test | Held-out records that pass the cleaning rules | Shows how much a clean test flatters each arm | Holdout IDs and pass rate by rule |
| Strata | Completeness, difficulty (handoffs, reopenings, length), customer type, outcome status | Gaps hide inside totals | Stratum key on every record |
| Metrics and uncertainty | Task metric per arm on R and K, per stratum, over several seeds, with bootstrap intervals over records | One run on one split can mislead | Score table with intervals and seed list |
Measure the gap by stratum before training anything
Start with counts, because they are cheap and often settle the question. For each stratum, divide its share in the clean subset by its share in the population sample; a ratio well below 1 marks a group the cleaning thinned out, and a stratum near zero cannot be rescued by reweighting.
The four strata cover most business records. Completeness counts filled required fields. Difficulty uses traces the system already keeps: reassignments, reopenings, thread length or elapsed time. Customer type uses account tier, size or segment. Outcome status separates completed work from cancelled, refunded, disputed and reopened cases.
Illustrative example: a support team's frame holds two years of tickets, and its export script keeps only tickets with every required field filled, one closing resolution and fewer than three reassignments. The shares below are invented to show the arithmetic and describe no real company.
Here the ratios answer the coverage question before any training run. A model trained on this clean subset would see roughly one refunded, cancelled or reopened case for every three or four it will meet in production, and it would rarely see the terse notes customers actually write.
| Stratum (illustrative) | Population sample share | Clean subset share | Ratio | What it means |
|---|---|---|---|---|
| All required fields filled | 58% | 100% | 1.7 | The rule itself; expected |
| Reassigned two or more times | 22% | 6% | 0.27 | Hard cases cut by almost three quarters |
| Small-account customers | 41% | 27% | 0.66 | A customer group under-represented |
| Refunded, cancelled or reopened | 18% | 5% | 0.28 | Most failure evidence gone |
| Notes under 20 words | 35% | 9% | 0.26 | Terse real-world input rarely seen |
How to read the cross-test results
Read arm C against arm P on both test sets; the pattern, not a single score, tells you what the cleaning did. Report each comparison with its interval and per stratum, and treat a difference smaller than the seed-to-seed spread as no difference.
At a glance
How to read the cross-test results
Two cautions keep the comparison honest. Size-match the arms, or a smaller clean arm will look worse for reasons unrelated to cleanliness. And keep R free of every cleaning rule except error correction, or the study quietly tests the cleaning against itself.
Checklist · 5 items
- C wins on the curated test K but loses on the representative holdout R: the cleaning bought polish on easy cases at the cost of coverage. Reverse the selection rules and keep the error fixes.
- C wins on both: the cleaning removed noise that hurt learning. Keep it, document it, and still report R, because buyers deploy on R-like data.
- P wins on both: the cleaning removed signal, usually ambiguity or failure evidence. Return to population sampling and use flags in place of filters.
- No difference on R while C scores higher on K: the clean test set flatters every model. Keep the training data and replace the test set with R.
- Either arm collapses on one stratum, such as reopened cases: that stratum needs its own supplement or an explicit exclusion in the dataset's claims, whatever the totals say.
0 of 5 ticked. Tick each item as you go.
Which cleaning steps to keep, flag or reverse
Keep steps that correct errors in place, flag steps that judge messiness, and reverse steps that drop whole records for traits linked to difficulty or outcome. The cross-test then decides which flags become filters for a specific use.
Open data projects already work this way. RedPajama-V2, an open pretraining corpus of over 100 billion documents from 84 Common Crawl snapshots, ships 30 billion of them with quality signals attached rather than as one filtered set, so each user picks filters for their own task. Datasheets for Datasets, the documentation framework by Gebru and colleagues, asks creators whether the raw data was saved in addition to the cleaned data, to support uses nobody anticipated. Near-duplicate removal deserves its own threshold experiment; treat it as an error fix only when copies are exact.
Checklist · 4 items
- Keep: fixing malformed dates, normalizing units and codes, merging exact duplicates, correcting labels a second reviewer confirms are wrong.
- Flag: missing fields, short or template-breaking notes, reassignment counts, low labeler agreement, outlier values. Store the flag and let each use decide.
- Reverse: dropping open, cancelled, refunded or reopened cases; dropping records with any missing field; dropping ambiguous labels without review; dropping underspecified requests from evaluation sets.
- Never: choosing which records to keep by their outcome when the model will predict or act on outcomes.
0 of 4 ticked. Tick each item as you go.
A sampling and documentation policy that preserves known limitations
Write the policy once and apply it to every delivery: the frame, the filters with their counts, an untouched holdout and both scores travel with the data. A buyer should be able to rebuild the clean subset from the frame and see exactly what it left out.
Bring in a statistician when strata are so thin that every interval overlaps, or when cleaning and outcome are so entangled that no size-matched arm can separate them. In those cases, narrow the dataset's claims rather than reporting a confident number.
- Keep the full pre-cleaning frame, with identifiers, stratum labels and dates, for as long as the dataset is in use.
- Log every rule that drops records, with its date, author and the count removed in each stratum.
- Hold out a population-sampled test set before any cleaning, and never select it by quality.
- Report scores on the representative holdout and the curated test side by side, per stratum.
- Mark ambiguous records with a label such as needs clarification instead of deleting them.
- State in the dataset card which uses the clean subset supports and which it does not, such as estimating failure rates.
Putting the representative holdout into a data request
Ask for the frame and a population-sampled holdout in the request itself, before any records are cleaned for delivery, and name the strata you will compare. A supplier who can produce the frame can also show you what each cleaning rule removed.
SourceX holds no ready-made datasets; it sources from U.S. businesses against each buyer's request. Next to what the request-data form already asks for (scope, volume, language, format, timing and licensing needs), state the acceptance test, the strata and the holdout rule from this page. SourceX takes delivery of nothing until the supplying company has signed its agreement on scope and terms it approved, so the sampling policy can be settled before preparation starts.
Frequently asked questions
Is clean data always better for training an AI model?
No. Correcting errors usually helps, but dropping whole records for being incomplete, long or ambiguous can remove the hard cases and failures a deployed model must handle. Test it: train size-matched clean and population-sampled arms and compare them on a representative holdout.
What is the difference between data cleaning and data filtering?
Cleaning changes values inside a record, such as fixing a date or normalizing a code, and leaves coverage intact. Filtering keeps or drops whole records by a rule. Filtering on traits linked to difficulty, customer type or outcome is where clean data sampling bias comes from.
How do you check a business dataset for coverage bias after cleaning?
Keep the pre-cleaning frame, draw a population sample from it and compare stratum shares with the cleaned set on completeness, difficulty, customer type and outcome status. Divide the clean share by the population share; ratios well below 1 show the groups the cleaning thinned out.
Should an evaluation set be cleaner than the training set?
Its labels should be checked more carefully, because label errors can change which model ranks first. Its selection should not be cleaner: an evaluation set that keeps only well-specified, complete cases overstates performance on real requests, so keep an ambiguous or underspecified slice.
Does removing mislabeled records cause sampling bias?
It can, when the filter cannot separate wrong labels from genuinely ambiguous cases, since both look uncertain to a model. Have a second reviewer confirm errors before dropping them, and mark ambiguous records instead of deleting them.
Sources
- NIST SP 1270 ties statistical and computational bias to systematic error from non-representative samples and notes it can occur without prejudice or discriminatory intent. Source
- Gururangan and colleagues (EMNLP 2022) found the GPT-3 quality filter favored newspapers from larger schools in wealthier, more educated, urban ZIP codes, and its judgments did not align with factuality or literary acclaim. Source
- SWE-bench Verified was introduced on August 13, 2024 as a subset of 500 problems that software engineers confirmed are solvable. Source
- OpenAI's annotators screened 1,699 random SWE-bench samples for underspecified issue descriptions and unit tests that reject valid solutions, producing a 500-sample verified subset. Source
- GDPval covers 1,320 tasks across 44 occupations, is limited to one-shot evaluation, and future versions will extend to more interactive, context-rich workflows. Source
- The GDPval paper reports a variant with shorter, under-specified prompts on which models performed worse, mainly because they struggled to work out the missing context. Source
- Northcutt, Athalye and Mueller estimated label errors averaging at least 3.3 percent across 10 popular test sets and showed such errors can change which model ranks first. Source
- Dataset Cartography found hard-to-learn examples often correspond to labeling errors, while training on ambiguous examples improved out-of-distribution performance with little or no in-distribution change. Source
- Anthropic advises that grader checks be inferable from the task description and describes a model that failed an evaluation as written while finding a better solution for the user. Source
- RedPajama-V2 holds over 100 billion documents from 84 Common Crawl snapshots, of which 30 billion come with quality signals rather than being pre-filtered. Source
- Datasheets for Datasets asks whether the raw data was saved in addition to the preprocessed, cleaned or labeled data, to support unanticipated future uses. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.