Alena from Sber: Why Evaluating an LLM on 20 Examples Is More Dangerous Than Not Evaluating It at All
The creator of Russian SuperGLUE and curator of the MERA benchmark at Sber describes the central paradox of industrial LLM evaluation: 10–20 test examples scored by an LLM judge look like quality control, but only produce false confidence. What teams need is a middle path: rigorous enough to support decisions, and light enough that people actually use it.
AI-processed from Habr AI; edited by Hamidun News
Alena, curator of the MERA benchmark and creator of Russian SuperGLUE, describes a systemic gap between academic rigor in evaluation and the real pace of development. She also explains why a "quick check" on 10 examples is more dangerous than having no check at all.
Academia versus industry
In academia, a benchmark is serious methodological work. It requires annotated datasets, justified metrics, data leakage checks, reproducible results and error analysis. A good test can take months and needs a dedicated team. But its results can be trusted.
Industry runs on a different clock. A team has to pick a model before the end of the sprint, test a new prompt version, compare two RAG pipelines, and find out whether quality dropped after an update, ideally by the next release rather than in six months. The academic approach simply does not fit that pace.
This gap produces two opposite scenarios. The first is minimal evaluation with no system behind it: a few examples before a demo, a quick eyeball of the answers, "seems to work." The second is the appearance of quality control: 10–20 queries, an LLM judge, an average score, a table dropped into a report.
Why 10 examples are more dangerous than zero
The author draws a counterintuitive conclusion: the second scenario is worse than the first. A standard "evaluation on twenty examples" looks like a process, but it builds false confidence on a statistically weak signal. The specific problems with this approach:
- The sample is too small: 10–20 examples do not give a reproducible result, and a different set will show different numbers
- An uncalibrated LLM judge is biased: it systematically prefers long, confident, well-structured answers, regardless of factual accuracy
- The average score hides failures: a model can get a high average score while failing completely on one particular type of task
- There is no baseline: without a fixed reference point, it is impossible to tell whether things got better or worse after a change
- Data leakage is not checked: the model may have seen the test examples during training, in which case the score measures nothing real
"The problem is that the second option often looks like quality control but is not.
It can be dangerous because it creates confidence where there is actually only a very weak signal."
Where this pattern comes from
Alena stresses that this is not about laziness or misunderstanding. Teams know perfectly well that evaluation matters. The classic academic approach is simply too heavy a tool for the real pace of development. What is needed is a middle path: rigorous enough to make decisions, and light enough that people actually do it.
The author is curator of the "MERA" project of the AI Alliance, a benchmark for evaluating Russian-language LLMs, and a contributor to the creation of Russian SuperGLUE and ruMTEB. Over five years of working with language models, she has seen the same gap in teams of every size, from startups to large corporations.
What this means
LLM evaluation is not a one-off procedure before a demo but production infrastructure. Teams that build it systematically save time diagnosing regressions and make fewer costly mistakes in production.
The article offers a practical guide for those stuck between academic rigor and a dangerous illusion of control.
- Apr 30, 2026Sber, Yandex and red_mad_robot show how AI is changing the developer's role
- May 17, 2026Sber study: how AI quietly substitutes its own assumptions for strategic decisions
- Nov 8, 2026Sber Open-Sources Universal Agent: Same Model, New Harness, 24 Points Better
- Jun 11, 2026OpenClaw fails phishing tests: AI agent can't spot forged emails
Key Questions Answered
What are the key details on Academia versus industry?
In academia, a benchmark is serious methodological work. It requires annotated datasets, justified metrics, data leakage checks, reproducible results and error analysis. A good test can take months and needs a dedicated team. But its results can be trusted. Industry runs on a different clock. A team has to pick a model before the end of the sprint, test a new prompt version, compare two RAG pipelines, and find out whether quality dropped after an update, ideally by the next release rather than in six months.
Why 10 examples are more dangerous than zero?
The author draws a counterintuitive conclusion: the second scenario is worse than the first. A standard "evaluation on twenty examples" looks like a process, but it builds false confidence on a statistically weak signal. The specific problems with this approach: The sample is too small: 10–20 examples do not give a reproducible result, and a different set will show different numbers An uncalibrated LLM judge is biased: it systematically prefers long, confident, well-structured answers,…
Where this pattern comes from?
Alena stresses that this is not about laziness or misunderstanding. Teams know perfectly well that evaluation matters. The classic academic approach is simply too heavy a tool for the real pace of development. What is needed is a middle path: rigorous enough to make decisions, and light enough that people actually do it. The author is curator of the "MERA" project of the AI Alliance, a benchmark for evaluating Russian-language LLMs, and a contributor to the creation of Russian SuperGLUE and ruMTEB.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.