nvidia-generative-ai-notes

Model Evaluation

Systematic evaluation is essential at every stage: after pretraining (does the model understand language?), after fine-tuning (does it follow instructions?), after alignment (is it safe and helpful?), and after quantization (did we lose accuracy?).

Automatic Metrics

Perplexity

Measures how well a model predicts the next token. Lower is better.

PPL = exp(-1/N Σ_{i=1}^{N} log P(x_i | x_{<i}))

• Use case: Language modeling quality, comparing model versions
• Benchmark datasets: WikiText-2, The Pile validation, C4
• Limitations: Doesn’t measure instruction-following, safety, or reasoning

BLEU (Bilingual Evaluation Understudy)

Measures n-gram overlap between generated text and reference:

BLEU = BP · exp(Σ_{n=1}^{4} w_n · log p_n)

Where p_n = modified n-gram precision, BP = brevity penalty, w_n = uniform weights (typically 1/4).

• Use case: Machine translation, summarization
• Range: 0-100 (higher is better)

ROUGE (Recall-Oriented Understudy for Gisting Evaluation)

Measures recall of n-grams from reference text:

• ROUGE-1: Unigram overlap
• ROUGE-2: Bigram overlap
• ROUGE-L: Longest common subsequence
• Use case: Summarization quality

Exact Match and F1

• Exact Match: Binary — does the prediction exactly match the reference?
• F1: Token-level precision/recall harmonic mean
• Use case: Extractive QA (SQuAD), closed-form answers

Pass@k

For code generation: generate k samples, check how many pass unit tests.

Pass@k = 1 - C(n-c, k) / C(n, k)

Where n = total samples, c = correct samples. Unbiased estimator of the probability that at least one of k samples is correct.

• Use case: Code generation (HumanEval, MBPP)

Benchmark Suites

Benchmark	Measures	Format	Tasks/Subjects
MMLU	World knowledge	Multiple choice	57 subjects (STEM, humanities, social sciences)
HellaSwag	Commonsense reasoning	Completion selection	10K scenarios
HumanEval	Code generation	Executable Python	164 programming problems
MBPP	Code generation	Executable Python	974 problems
GSM8K	Math reasoning	Free-form numerical	8.5K grade school math
MATH	Advanced math	Free-form	Competition-level problems
TruthfulQA	Factual accuracy	QA pairs	817 questions designed to elicit falsehoods
ToxiGen	Safety and bias	Classification	274K toxic/benign statements
ARC	Science reasoning	Multiple choice	7.7K grade-school science questions
WinoGrande	Coreference resolution	Fill-in-the-blank	44K pronoun resolution

LLM-as-Judge Evaluation

Use a strong LLM (e.g., GPT-4) to evaluate model outputs:

• MT-Bench: Multi-turn conversation quality scored 1-10
• AlpacaEval: Instruction-following win rate vs. reference model
• Chatbot Arena: Crowdsourced pairwise comparisons (Elo rating)

Advantages: captures nuance that automatic metrics miss. Limitations: expensive, potential bias toward certain response styles.

Evaluation Frameworks

EleutherAI lm-evaluation-harness

The standard open-source evaluation framework:

• 200+ tasks pre-implemented
• Zero-shot and few-shot evaluation
• HuggingFace model support, extensible to custom formats
• Community-maintained benchmarks

lm_eval --model hf --model_args pretrained=meta-llama/Llama-3-70B \
    --tasks mmlu,hellaswag,gsm8k \
    --num_fewshot 5 \
    --batch_size 8

HELM (Holistic Evaluation of Language Models)

Stanford’s comprehensive evaluation framework:

• Evaluates across 7 metrics: accuracy, calibration, robustness, fairness, bias, toxicity, efficiency
• 42 scenarios covering diverse use cases
• Standardized reporting format

OpenAI Evals

• Custom evaluation definition format
• Support for LLM-as-judge patterns
• Composable evaluation functions

Evaluation at Each Stage

Stage	Key Metrics	What to Watch
Pretraining	Perplexity, MMLU, HellaSwag	Training loss convergence, benchmark trends
SFT	MT-Bench, instruction-following accuracy	Overfitting to training format
Alignment	TruthfulQA, ToxiGen, MT-Bench	Alignment tax (capability regression)
Quantization	Perplexity delta, MMLU delta	<1% degradation on critical metrics
Deployment	Latency, throughput, user satisfaction	A/B test against baseline

Best Practices

• Multiple metrics: No single metric captures model quality — use a diverse evaluation suite
• Hold-out data: Never evaluate on training distribution
• Track across stages: Compare metrics after each training/optimization stage
• Domain-specific benchmarks: Add evaluations specific to your production use case
• Human evaluation: Automated metrics supplement but don’t replace human judgment for subjective qualities
• Regression testing: Maintain a fixed evaluation set to detect quality regressions across model versions