← TinyGPT · docs · devlog · roadmap · speedup
source: docs/learn/advanced-ml-systems-eval.md · view on GitHub ↗

Advanced architecture, RAG/agents, eval & system design

The breadth half of staff interviews. Format: what’s probed, best source, in the repo. Transformer basics (attention mechanics, gradient descent) live in the session-0X series + llm-mechanics-fundamentals.md; this doc is the senior-depth layer on top.

Modern architecture

1. The 2025-era decoder block. Defend each deviation from the 2017 paper: pre-norm vs post-norm (gradient stability at depth), RMSNorm (drops mean-centering, cheaper, same quality), SwiGLU gated FFN. “Walk me through a modern decoder layer and justify each change.” Learn: Transformer Design Guide · senior In repo: the Qwen3 block in TinyGPTModel; mechanics in llm-mechanics-fundamentals.md.

2. RoPE. Rotary embeddings encode relative position via complex-plane rotation of Q/K, parameter-free; extension via NTK/YaRN frequency interp (see advanced-llm-inference.md §14). “Why does RoPE generalize to longer context than learned absolute positions?” Learn: RoFormer · senior/staff

3. Attention as matmuls — the whiteboard. Live derivation: X∈[n,d] → Q,K,V via [d,d_k] projections; scores QKᵀ∈[n,n]; row-softmax; out [n,d_v]. Must explain the 1/√d_k factor: dot products of unit-variance vectors have variance d_k, so scaling renormalizes to keep softmax out of the saturated/vanishing-gradient regime. Expect FLOPs + O(n²) memory questions. This is the “relate linear algebra to transformers” round. Learn: The Annotated Transformer · senior In repo: session-06-tokenization-embeddings.md + llm-mechanics-fundamentals.md build this ground-up.

4. Why decoder-only won. Causal masking enables the KV-cache and cheap autoregressive serving; in-context learning + scaling. “When would you still reach for an encoder (BERT/embeddings)?” Learn: Decoder-Only Workhorse · senior In repo: model-vs-agent.md; session-04-ml-paradigms.md on encoder–decoder.

(MoE routing → advanced-llm-training.md §7; GQA/MLA + KV economics → advanced-llm-inference.md §12.)

RAG & agents

5. RAG retrieval design. Chunking (fixed/semantic/recursive, overlap) as a product choice; hybrid BM25+dense (sparse catches exact IDs/SKUs, dense catches paraphrase). “Rare-SKU queries fail — fix retrieval.” Learn: Hybrid Search · senior In repo: Pace PaceLocalRetrieval is BM25 lexical + best-effort embedding re-ranker today (docs/prds/local-rag-layer.md).

6. Rerankers & embeddings. Bi-encoder (fast ANN first stage) vs cross-encoder (accurate rerank of top-k); when to fine-tune embeddings; MTEB for selection. “Recall fine, top-3 precision poor — what stage?” Learn: Retrieve & Re-Rank · senior In repo: tinygpt rerank-train / rerank-eval / eval-mteb; mxbai-embed.

7. RAG evaluation. Measure retrieval (recall@k, MRR/nDCG) separately from generation faithfulness/groundedness; knowledge conflict (parametric vs retrieved). “Answers wrong but retrieval looks right — localize it.” Learn: RAGAS · senior/staff

8. When NOT to use RAG / query rewriting. Adaptive transforms (multi-query, HyDE, decomposition) on weak retrieval; skip RAG for small stable corpora (fine-tune / long-context) or low-latency paths — the RAG-vs-long-context staff debate. Learn: Contextual Retrieval · staff

9. Agents: ReAct, planning, memory, tools. Agent = LLM + planning + memory + tools; reason↔act loop; short-term (context) vs long-term (vector) memory; structured function-calling + error recovery + loop bounding. “Design a travel-booking agent — where does it loop/fail, how do you bound it?” Learn: Lil’Log: LLM Agents · senior In repo: tinygpt agent; Pace’s plan-act-observe loop; agent-context-hierarchy.md.

10. Agent evaluation (trace-based). Score tool-call correctness, planning quality, task completion via step-level traces — not just the final answer. “Eval an agent with 20 tool calls when only the last output shows.” Learn: Agent Eval Guide · staff In repo: B23 agent-eval-protocol (pass@1 over repeated runs ± σ); tinygpt eval-bfcl / eval-tau-bench — and today’s A1 run is this end-to-end.

Evaluation depth

11. LLM-as-judge + its failure modes. The real signal is naming biases: position, length/verbosity, self-preference, concreteness, prompt-injection. Mitigate: pairwise + order randomization, rubric + CoT, validate vs human labels. “Your judge prefers longer answers — prove it and correct it.” Learn: Hamel: Evals FAQ · senior/staff In repo: tinygpt judge (E7, JudgeShim.swift); strict-scorer mode.

12. Perplexity — definition, uses, limits. PPL = exp(mean per-token NLL) = exp(cross-entropy). The catch: comparable only under the same tokenizer, measures next-token confidence not correctness, lower PPL ≠ better generations. “When is perplexity meaningful vs misleading?” Learn: HF Perplexity · senior In repo: training loss in Train.swift is cross-entropy (exp it for PPL); the eval suite (eval-matrix-2026-06-08.md) is the “PPL isn’t enough → use task evals” argument made concrete.

13. Contamination, pass@k, calibration. Test leakage inflates scores — detect/guard via held-out/private sets, canary strings, decontamination; pass@k for code (functional correctness over k samples); calibration (ECE, reliability diagrams). Learn: Contamination Survey · staff In repo: mac-assistant-judgment benchmark ships a real contamination check (Jaccard ≥0.6 vs train); tinygpt eval-humaneval is pass@k.

ML system design & classic-ML depth

14. Recommender / ranking / feed design. The retrieval → ranking → re-ranking funnel; two-tower/ANN candidate gen vs heavy ranker; objectives (engagement vs relevance vs freshness vs diversity); cold-start. The most-asked staff system-design prompt. Learn: Eugene Yan: System Design for Discovery · staff

15. Training-serving skew & feature stores. Skew from features computed differently train vs serve; the fix is one feature store + point-in-time joins (no label leakage). “Offline AUC great, online tanked — debug.” Learn: Rules of ML · senior/staff

16. A/B testing & guardrail metrics. Goal vs guardrail metrics, power/MDE/sample size, novelty effects, peeking/multiple-comparison corrections, shadow→canary→rollback. “Goal up but a guardrail moved — ship?” Learn: Microsoft ExP: Metric Pitfalls · senior/staff

17. Classic-ML senior depth. Bias-variance as capacity lever; L1/L2 (sparsity vs shrinkage); calibration (Platt/isotonic — critical for credit risk); causal inference / uplift (two-model / meta-learners, propensity, DiD, doubly-robust) under class imbalance; imbalance handling (class weights, PR-AUC over accuracy). Maps directly to a credit-risk background. Learn: Elements of Statistical Learning · senior/staff In repo: the SAE-vs-PCA contrast (speech-and-systems-topics.md §5) is the dimensionality-reduction version of this.

Suggested order

3 and 11–13 are the highest-differentiation for this profile (the linear-algebra whiteboard + judge-bias enumeration + calibration). 14–17 are pure system-design/DS study; ESL is the slow material — budget for it.