llm-evaluation

# LLM Evaluation (Deep Workflow) Evaluation turns “it feels better” into **reproducible evidence**. Design around **failure modes** your product cares about—not only aggregate scores. ## When to Offer This Workflow **Trigger conditions:** - Prompt or model change; need **before/after** proof - Building **CI** for LLM outputs; flaky quality in production - RAG/agents: **grounding**, **tool use**, **safety** regressions **Initial offer:** Use **six stages**: (1) define quality & constraints, (2) build datasets & rubrics, (3) automatic metrics, (4) human evaluation, (5) regression & gates, (6) online validation & iteration. Confirm **latency/cost** budgets and **risk** (PII, safety). --- ## Stage 1: Define Quality & Constraints **Goal:** Name **dimensions** that map to user harm if they fail. ### Typical dimensions (pick what matters) - **Correctness** / task success; **groundedness** (RAG); **faithfulness** to sources - **Safety**: policy violations, jailbreaks, PII leakage - **Style**: tone, brevity, format (when product-critical) - **Robustness**: paraphrase, multilingual, edge inputs ### Constraints - Max **tokens**, **latency** p95, **cost** per request; **locale** requirements **Exit condition:** Weighted **priority** of dimensions; **non-goals** stated. --- ## Stage 2: Datasets & Rubrics **Goal:** **Fixed** eval sets + **clear** scoring rules. ### Practices - **Stratify** by intent: easy/medium/hard; **adversarial** slice separate - **Rubrics**: 1–5 scales with **anchors**; **binary** checks for safety - **Version** datasets (git or table); **no** silent edits without changelog - **Privacy**: synthetic or **redacted** real examples per policy **Exit condition:** **Golden set** size justified; **inter-rater** plan if human scoring. --- ## Stage 3: Automatic Metrics **Goal:** **Fast** signals—know **limitations**. ### Options - **Reference-based**: BLEU/ROUGE—often weak for assistants - **Model-as-judge**: fast, biased—**calibrate** vs human - **Task-specific**: exact match, JSON schema validity, tool-call args match - **RAG**: citation overlap, **nugget** recall, entailment models (use carefully) ### Hygiene - **No** training on test; **detect** **leakage** from prompts **Exit condition:** Each auto metric has **known blind spots** documented. --- ## Stage 4: Human Evaluation **Goal:** **Authoritative** judgment where automatic metrics lie. ### Design - **Sample size** for confidence; **blind** A/B when possible - **Guidelines** + **examples**; **adjudication** for disagreements - **Locale-native** raters when language quality matters **Exit condition:** **Human** scores correlate **enough** with auto for ongoing monitoring—or you rely on human for release. --- ## Stage 5: Regression & Gates **Goal:** **Block** bad deploys in **CI** or **release** pipeline. ### Gates - **Must-pass** suites: safety, critical user journeys - **Trend** tracking: **not** only point-in-time - **Canary** with **online** metrics (see Stage 6) ### Artifacts - **Report**: model/prompt id, dataset versions, scores, **diff** **Exit condition:** **Rollback** criteria defined before rollout. --- ## Stage 6: Online Validation **Goal:** **Production** truth—shadow, A/B, or gradual ramp. ### Signals - **Implicit**: thumbs, edits, task completion, support tickets - **Explicit**: user ratings (sparse) ### Causality - **Confounds**: seasonality, cohort—**control** where possible --- ## Final Review Checklist - [ ] Quality dimensions prioritized for the product - [ ] Versioned eval sets and rubrics - [ ] Auto + human roles explicit; limitations documented - [ ] Release gates and rollback tied to metrics - [ ] Plan for online feedback loop ## Tips for Effective Guidance - **Slice** metrics—averages hide **regressions** on critical intents. - For **agents**, evaluate **trajectories**, not only final text. - Never claim **objective** truth—evaluation is **operationalized** judgment. ## Handling Deviations - **No labels**: start with **smallest** **pairwise** comparison set + **spot** human review. - **High-stakes** (medical/legal): **human-in-the-loop** gate; disclaim **limits** of auto eval.