Most "Model Problems" Are Data Problems
When accuracy looks magical, eval looks noisy, or an LLM "forgets" instructions, inspect the data path before the architecture. The mistakes below show up in startups and enterprises alike—usually inside a CSV someone trusted too much.
Mistake: Target Leakage
Features include information unavailable at prediction time—future statuses, post-outcome notes, or IDs that encode the label.
- Symptom: unrealistically high validation scores.
- Fix: timeline audits; drop post-event columns; document feature availability.
Mistake: Train / Eval Contamination
Near-duplicate prompts or rows appear in both train and eval, or eval is "cleaned" until scores look good.
- Symptom: demos wow; production fails.
- Fix: freeze eval, dedupe across splits, ban silent relabeling of eval.
Mistake: Inconsistent Labels
positive, Positive, pos, and + all mean different strings to a trainer.
- Symptom: classes explode; metrics oscillate.
- Fix: enum map, labeling guide, validation lint in CI.
Mistake: Excel Silent Corruption
Opening CSV in Excel turns IDs into scientific notation, parses SEPT1 as a date, and re-saves with a new delimiter.
- Symptom: join failures; weird categories.
- Fix: import via Data → From Text/CSV; keep string IDs; use Excel to CSV UTF-8 deliberately; preview in Online CSV Editor.
Mistake: Messy File Mechanics
Wrong encoding, mixed delimiters, or two header rows.
| Symptom | Likely cause |
|---|---|
| Mojibake names | Non-UTF-8 encoding |
| One fat column | Wrong delimiter |
| Shifted fields | Unescaped quotes/newlines |
| Mystery first column name | UTF-8 BOM |
Mistake: Shipping PII to External AI
Pasting customer exports into consumer chat tools for "a quick clean."
- Symptom: policy and trust failures—not just bad models.
- Fix: redact, synthesize, or use approved enterprise workspaces; clean locally first.
Mistake: No Grain Definition
Rows mix orders and order lines, or customers and events, without saying so.
- Symptom: duplicated metrics; impossible joins.
- Fix: write "one row means…" at the top of the data card; split files by grain.
Mistake: Trusting LLM Extractions Tables Blindly
Models invent cells when extraction is hard. Unvalidated JSON/CSV becomes training fuel.
- Symptom: structured pipelines full of plausible fiction.
- Fix: schema validation, spot checks, confidence thresholds, human review on critical fields.
Diagnostic Checklist
Run this when results smell wrong.
- Compare row counts raw → clean → train/eval.
- Profile nulls, dupes, and label distribution.
- Scan for leakage columns.
- Re-open CSV without Excel defaults.
- Recompute a metric by hand on 20 rows.
Real-World Examples
Postmortems in miniature.
The 0.99 AUC
A "payment_cleared_at" feature leaked outcomes. Removing it dropped AUC to a believable range.
The prompt that worked once
Eval prompts were edited in the same Sheet as train examples; contamination inflated win rates.
The vanished leading zeros
Store IDs lost zeros in Excel; store-level features collapsed. String dtype fixed training joins.
Best Practices
Prevention over heroics.
- Data contracts and labeling guides.
- Frozen eval sets.
- Scripted cleaning; versioned artifacts.
- Redaction before external AI.
- Convert and preview with Convert CSV Online before trust.
Why Use Convert CSV Online?
Convert CSV Online is free, browser-based, and requires no account for everyday conversions. Catch delimiter, header, and Excel export issues in the Online CSV Editor before they become "AI bugs." Client-side workflows work on Windows, macOS, and Linux browsers.
Verify the artifact
If the table looks wrong to you, it will look wrong to the model—fix the file, not only the prompt.
Conclusion
Common AI data mistakes—leakage, contamination, label chaos, Excel corruption, and blind trust in extracted tables—are avoidable with contracts, frozen evals, and disciplined CSV hygiene.
FAQ
What is the most common AI data mistake?
Target leakage and train/eval contamination are among the most damaging—both create misleadingly strong metrics.
How do I detect data leakage?
Audit when each feature is known relative to the prediction moment, and be suspicious of near-perfect offline scores.
Why do Excel edits break AI datasets?
Excel auto-converts types, dates, and delimiters. Re-export carefully or avoid Excel for golden training files.
How do I fix inconsistent labels?
Define an enum, map synonyms, lint the label column in CI, and update the labeling guide.
Can I paste production CSV into ChatGPT to clean it?
Not if it contains PII or secrets. Redact or clean locally with approved tools first.
How should I validate LLM-extracted CSV?
Check schema, required fields, and spot-check rows against source documents before training or loading a database.
References
Convert your CSV in the browser
Preview, clean, and convert CSV files free with Convert CSV Online—no installation and no account required for everyday conversions.