Evaluation
25 bites tagged Evaluation — interview questions with model answers, and 60-second explainers.
Designing an LLM red-teaming framework
Taxonomy of harms, automated adversarial prompt generation via attacker models and mutation, a classifier to triage outputs, and severity-by-likelihood prioritization. systematic safety probing.
The alignment tax and capability trade-offs
Alignment tax is capability lost from safety tuning, measured as benchmark or task-success deltas before and after; a product decision weighs over-refusal against harm risk. cost of safety interventions.
Red teaming a generative model
Deliberately probe for harmful outputs across categories, document jailbreaks, and automate with adversarial prompt generators plus classifier-based judging. Adversarial safety evaluation.
Evaluating faithfulness and compositionality in multimodal models
Use targeted probes with hard negatives, attribute-relation binding tests, and structured grounding checks; note BLEU rewards surface overlap not correctness. Going beyond n-gram metrics.
Detecting catastrophic forgetting in continual fine-tuning
Maintain a frozen held-out benchmark of original capabilities, evaluate after every fine-tune, track per-capability deltas, and alert on regressions. guarding original skills during continual training.
Evaluating a multi-tool LLM agent
Measure end-to-end task success, plus trajectory quality like correct tool choice and arguments, efficiency via steps and cost, and robustness to errors and edge cases. evaluating multi-step, tool-using behavior.
Designing a production LLM summarization eval
A representative gold set, quality via human or LLM-as-judge plus faithfulness checks, and operational metrics like p95 latency and cost per request. building task-specific eval beyond leaderboards.
Detecting RAG hallucinations with a confidence score
Decompose the answer into claims, verify each against retrieved context with NLI or an LLM judge, aggregate into a faithfulness confidence score, and flag unsupported claims. groundedness verification design.
Setting up an LLM-as-a-judge evaluation
Define rubric, prefer pairwise comparison, randomize order, calibrate against humans; control position, verbosity, and self-preference bias. rigorous LLM-as-judge design.
Reference-free evaluation for open-ended dialogue
ROUGE punishes valid paraphrases; use reference-free LLM-as-judge or learned scorers rating coherence, relevance, and groundedness. knowing why n-gram overlap fails for dialogue. defending ROUGE for open-ended generation.
How MMLU works and the contamination problem
MMLU is multiple-choice across 57 subjects scored by accuracy; contamination means test items leaked into pretraining, inflating scores. understanding of LLM benchmarks and their flaws.
How to evaluate a RAG system end to end
Measure retrieval (recall, precision, MRR, NDCG) and generation (faithfulness, answer relevance) separately, plus end-to-end correctness. ability to decompose RAG evaluation. judging only final answer quality with one number.
How FID is calculated versus Inception Score
FID fits Gaussians to Inception features of real and fake images then measures Frechet distance; it uses real references and detects mode collapse. deep grasp of generative metrics. claiming IS uses real data.
Standard metric for image generation quality
Name FID, explain it compares feature distributions of real and generated images via a pretrained network. knowledge of generative model evaluation. relying only on eyeballing samples or pixel-level MSE.
Why human evaluation is the gold standard
Humans judge fluency, helpfulness, and correctness that n-gram or distribution metrics miss; automated scores correlate weakly with quality, are gameable, and penalize valid diverse outputs. limits of automated metrics.
Perplexity versus BLEU for LMs
Perplexity measures intrinsic next-token prediction quality needing no references; BLEU measures n-gram overlap with reference outputs for tasks like translation. choosing the right LM metric.
Evaluating image generation: FID and IS
FID compares feature distributions of real and generated images, lower is better; Inception Score rewards confident, diverse classes but ignores real data. knowledge of generative image metrics.
Evaluating a RAG system end to end
Measure retrieval with context recall or precision, and generation with faithfulness and answer relevance, attributing failures to the right stage. ability to separate retrieval and generation quality.
What does N-way K-shot classification mean?
N is classes per episode, K is labeled examples per class in the support set, prediction is on a separate query set. few-shot evaluation vocabulary. confusing K with total training data or swapping N and K.
How is IoU computed and why prefer mIoU?
IoU is intersection over union of predicted and true pixels; mIoU averages per class; pixel accuracy is dominated by background. segmentation metrics under imbalance. equating accuracy with IoU.
Evaluating generative models with FID versus IS
FID compares Inception feature distributions of real and fake images via Frechet distance between two Gaussians; it uses real data as reference and detects diversity issues, unlike IS which uses no real… rigorous generative evaluation.
Hallucination Detection in LLMs
Hallucination detection is the set of techniques for flagging when a language model states something fluent but false or unsupported, using signals like self-consistency, model uncertainty, and grounding against retrieved evidence to catch fabrications before…
Evaluating RAG Systems: Metrics for Retrieval and Generation
Evaluating a RAG system means grading its two parts: retrieval and generation. Metrics like relevance check if the right documents were found, while faithfulness and accuracy check if the final answer correctly uses those documents.
LLM-as-a-Judge: Using Models to Grade Models
Instead of paying humans to rate AI outputs, LLM-as-a-Judge uses a powerful 'judge' model to do it automatically. This is used to evaluate chatbot responses or summarization quality, but the main footgun is assuming the judge model is unbiased or perfectly…
Get Evaluation bites daily.
Five a day, five minutes, offline. With quizzes so it sticks.
Open testing — you’ll join as an early tester.