How to read benchmark scores when choosing a small model

Treat a benchmark score as a filter, not a decision. Ask what it measures, who ran it, whether the test set leaked into training, and whether it resembles your task. Then run your own eval on real inputs before choosing a model.

Six questions to ask

1. What does it measure? MMLU tests broad knowledge. MT-bench scores chat quality. HumanEval and MBPP measure code generation. None measure your support tickets.
2. Who ran it? The Phi-3 report bases its claims on academic benchmarks and internal testing. That is normal and useful, but internal numbers are the authors' own and may use their own settings.
3. Which settings? Prompts, number of examples and sampling affect scores. Scores from different papers are not always comparable.
4. Could the test leak into training? Public benchmark questions spread across the web, and web data feeds training. Scores may be inflated.
5. Is the metric gameable? A model tuned to the benchmark can look better than it is.
6. Does it test what hurts? Rare cases, refusals, long inputs and safety are often missing.

Examples from published reports

The Phi-3 report says phi-3-mini scores 69 percent on MMLU and 8.38 on MT-bench, and describes it as rivalling Mixtral 8x7B and GPT-3.5. The same family's phi-3-small and phi-3-medium score 75 and 78 percent on MMLU. The Llama 2 report describes fine-tuned chat models from 7B to 70B parameters and says human evaluations of helpfulness and safety suggest Llama 2-Chat may be a suitable substitute for closed-source models. These are the authors' claims in their own reports.

Benchmarks can mislead on truth

The TruthfulQA paper is a useful caution. It has 817 questions across 38 categories, built so that some humans would answer falsely because of misconceptions. The authors report that larger models were generally less truthful, with the best model scoring 58 percent truthful against 94 percent for humans. Bigger or higher on a general score does not mean better on every property.

Researchers also doubt chatbot leaderboards

The QLoRA authors argue that current chatbot benchmarks are not trustworthy for measuring chatbot performance levels. Take that as one view, and one more reason to run your own tests.

A quick procedure

1. Shortlist two or three models using public scores.
2. Build 100 to 300 real examples, see write evals before you fine-tune.
3. Run all candidates with the same prompt and settings.
4. Score by category and note cost and latency.
5. Re-run when a provider updates a model version.

What to record

Model name and version, date, prompt, settings and scores. Without these a result cannot be reproduced.

Quick check

Sources

• Phi-3 Technical Report, fetched 19 October 2026.
• Llama 2: Open Foundation and Fine-Tuned Chat Models, fetched 19 October 2026.
• TruthfulQA: Measuring How Models Mimic Human Falsehoods, fetched 19 October 2026.
• QLoRA: Efficient Finetuning of Quantized LLMs, fetched 19 October 2026.

Last reviewed: 19 October 2026.