The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
NaHero’s reported experience shows why matching labels across translated exam questions is a separate problem from producing plausible labels. The builder of the cloud-certification practice platform says independent first-pass tagging of its Portuguese and English versions agreed about 73% of the time on Google Workspace Admin. The reported response was to clarify ambiguous exam domains and use one language’s tags as the source of truth for its translation. That can make paired labels consistent; it does not prove they are correct.
The account comes from NaHero builder Mateus in a DEV Community article published October 6, 2026. Its implementation details and figures are the author’s report, not independently verified results, and the article does not provide a human-verified ground truth or an accuracy score.
What NaHero was tagging
NaHero is described as a timed-practice-exam product for AWS, Azure, and Google Cloud certifications, with exams in Brazilian Portuguese and English. Mateus reports a bank of about 8,700 questions across 30 exams, with each exam available in both languages. The labels assign questions to the exam’s domains so the product can report results by subject area and help users decide what to study.
That use makes a translation mismatch more than a cosmetic issue. If equivalent Portuguese and English questions receive different domain labels, users can get different per-domain results and, potentially, different study guidance depending on the language of the attempt.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
How the first-pass tagging worked
Closed-set labels, generated in batches
The builder says he used claude -p, Claude Code’s non-interactive mode, to label questions in batches of 40. Each prompt included the domain list for that particular exam. The generated labels were checked against that fixed list instead of accepting arbitrary new domain names. Mateus estimates the full bank took about 220 calls.
Constraining outputs to an exam-specific list helps prevent a model from inventing labels or varying their spelling. It does not resolve every boundary case: a question may plausibly fit more than one domain, or a domain name may be too broad to guide consistent classification.
Consistency was checked across languages
Mateus says he tagged the Portuguese and English versions independently and compared the results. For Google Workspace Admin, the first-pass labels agreed about 73% of the time. He regarded the disagreements as a product defect because domain-level performance feeds into the user’s next study choices.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
This is an agreement figure for one reported exam, not an accuracy figure. Two independent passes can agree and both be wrong; they can also disagree where the domain boundaries are genuinely ambiguous. The article does not describe a human-labeled reference set or report how often either language’s first-pass label was correct.
What changed after the disagreement
1. Add exam-specific disambiguation
The first intervention was to add lines clarifying domain meanings in the prompt. The example Mateus gives is “billing and pricing.” Exam-specific guidance can help distinguish nearby categories, especially when a short domain label is open to more than one interpretation.
This change addresses ambiguity in the labeling instructions. The account does not report a post-change agreement rate or an independent accuracy result, so it does not establish how much the clarification improved either measure.
2. Choose one language as the label source
The second intervention was to tag the English version and copy those labels to the Portuguese counterpart, rather than classify both versions independently. Copying labels makes paired questions match by construction and avoids language-specific differences in model interpretation.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsThat is a consistency mechanism, not a correctness check. If the source-language label is wrong, copying it propagates the error. Mateus makes the caveat explicit: “The first change still matters, because agreement alone doesn’t make the English tag correct.”
How to interpret the two interventions
| Intervention | What it addresses | What it does not establish |
|---|---|---|
| Clarify ambiguous domains in the prompt | Reduces room for different interpretations of exam-specific labels. | The reported account gives no measured post-change agreement or correctness score. |
| Copy labels from one language to its translation | Ensures the paired versions share a label. | Does not independently verify that the shared label is right. |
The useful distinction is between consistency and accuracy. A system can optimize for matching labels across translations, but correctness still depends on whether the selected label fits the question. The account supports the use of a single source of truth for paired labels; it does not show that English is inherently the more reliable source or that this approach improves accuracy.
Rank #4
How study plans used the labels
The reported study-plan pipeline separated numeric computation from generated prose. SQL calculated hits per domain, missed questions, and prior scores. DeepSeek V4.1 Flash then received those values and returned JSON text for the plan.
The product checked that returned domain names matched the supplied input and disallowed links. According to Mateus, it stored a plan once per attempt, including failures, to avoid repeatedly generating a plan for the same attempt. This design keeps the model from being responsible for calculating the underlying performance figures, while still requiring validation of its generated output.
What the reported product figures do—and do not—show
Mateus’s article gives a snapshot of NaHero as of October 6, 2026. These are project-reported figures for one product and period, not benchmarks for exam-preparation services, LLM labeling, or customer acquisition.
Best Value
| Reported measure | NaHero figure and qualification |
|---|---|
| Registered users | About 95, with about 75% from Brazil. |
| New signups | About 30 per week, with no ads running at the time described. |
| CLF practice scores | Most reported attempts scored between 23 and 47 out of 65; the article says passing is around 46. |
| Second-exam use | About 15% of users took a second exam in their first week. |
| Google Ads test | R$100 spent for 71 clicks; reported cost per click ranged from R$0.31 to R$0.90, click-to-signup conversion was 11%, and one annual subscriber resulted. |
| LinkedIn posts | The author says they brought in three times what Google Ads did; the article does not state the comparison’s measurement method. |
| Premium plan | Reported at R$10 per month or R$90 per year, adding question explanations, a study plan, and practice questions on the user’s weakest domain. |
The article describes all 30 exams, unlimited attempts, scores, per-domain results, and the ability to resume an attempt as free. These details give context for the product’s use of domain labels, but they do not validate the tagging method.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Practical lessons for cross-language classification
- Define the label set for the specific task and validate model output against it; a fixed vocabulary limits label drift but does not settle ambiguous cases.
- Compare equivalent items across languages when independent tagging could affect downstream results, and treat disagreement as a signal to inspect definitions and edge cases.
- Keep agreement separate from accuracy. If correctness matters, evaluate against reviewed examples or another defensible reference, not agreement alone.
- If you copy labels from one language to another, document that as a consistency choice and retain a way to catch errors in the source labels.
- Keep calculations deterministic where possible, then validate generated output against the data and schema the model was given.
NaHero’s reported case is a concrete example of a translation consistency problem in a product where labels influence study guidance. It suggests a workable division of labor—clarify the taxonomy, establish one label source for translations, and validate generated study-plan output—while leaving the central accuracy question unanswered by the figures reported.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
