Classifier hits 96% on test set, but duplicate emails exposed a leak
11/16/2026 — 11/16, 15:19·1 sources (all)·1 reports (all)
Key takeaways
- A team trained a support-email classifier on 40,000 labeled messages and got 96% on an 80/20 split, but only about three-quarters were route…
Background & analysis
On November 16, 2026, the AI section of DEV Community published a retrospective account of a support-email routing model that looked far more accurate on paper than it was in production. A team trained a small model to sort support emails into categories such as Delivery, Billing, Damaged goods, and Account changes, using 40,000 labelled emails. After shuffling the data, the team split it eighty-to-twenty, assigning 80 percent to training and 20 percent to testing. The model posted a test score of 96 percent, and the team shipped it without further scrutiny.
In the first month after deployment, however, the model routed only about three out of four emails correctly, a clear gap from the reported test result despite using the same model and the same class of emails. When the authors finally examined the data, they found that the same batch of duplicate emails had leaked into both the training set and the test set. Because of this overlap, the test set was not answering the question the team had intended. Instead, it was answering a different question, which meant the 96 percent figure never described the real-world task. The article does not describe any corrective steps taken after the discovery, leaving the account at the point where the data flaw had been identified.
AI-generated from 1 reports · updated 21 hours ago
- Heat index
- 54
- All sources
- 1
- All reports
- 1
- First seen
- 22 hours ago
Reports on this story headlines open the original
A team trained a support-email classifier on 40,000 labeled emails and saw 96% on a held-out split, yet only about three in four emails were routed correctly in the first month. The author traced the gap to duplicate emails sitting in both train and test sets, so the test was answering a different question than the one intended.
DEV Community · AIAI score 54
Other stories people are talking about
- Heat index 1345RisingNvidia board not discussing Jensen Huang's return, probe into his conduct advances22 sources
- Heat index 1241SurgeCalifornia Sues OpenAI Over Agents Escaping Tests, Intruding into Hugging Face15 sources
- Heat index 901Alibaba releases MoE model Qwen4-Max with 1M token context for low-cost, long-running tasks17 sources
- Heat index 695US Commerce Department issues interim final rule restricting remote access to AI chips via overseas data centers11 sources
- Heat index 532Anthropic to begin trading on Nasdaq this week under ticker ANT9 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

