Explore Prompts

Build an eval harness: recall@k, calibrated precision, answer faithfulness, and human-time-to-verify. Include topic-aware test buckets and data drift alarms.

Architect a retrieval stack with hybrid search, temporal decay, dedup, and passage-level citation anchors. Define fact-grounding checks and failure messages; include freshness reindex cadence.

Propose a self-play generation strategy where a teacher model drafts, a critic model scores, and a curator enforces diversity/novelty. Provide leakage and drift monitors.

Design a pipeline to mine high-quality instructions/solutions from forums, docs, and code. Include classifier-based filtering, self-checking, and multilingual normalization.

Write a data governance spec: license screening, PII scrubbing, near-duplicate collapse, contamination checks vs eval sets, and audit trails. Provide rejection reasons and exception handling.

Specify a training stack with SFT on curated data, preference optimization (DPO/ORPO), and optional RLHF. Include reward hacking tests, guardrails, and evals that predict production behavior.

Compare FlashAttention-3, RingAttention, Multi-Query, and Hybrid-Selective attention for 128k+ contexts. Provide kernel-level trade-offs, memory footprints, and document a migration path.

Engineer a KV-cache offload strategy spanning HBM→HBM2e→CPU RAM→NVMe. Define admission/eviction, compression, and reuse heuristics; simulate hit rates across context lengths (8k–256k).

Create a quantization evaluation suite (GPTQ/AWQ/RTN): perplexity, zero-shot accuracy, calibration set selection, and layer-wise sensitivity. Output deployment guidelines by architecture and hardware target.

Design an expert-parallel MoE serving topology: gate calibration, capacity factor, expert sharding, and interconnect constraints (NVLink/IB). Provide hot-spot diagnostics and expert-drop policies for brownout resilience.

Propose an implementation of speculative decoding with a 1–3B drafter and 13–34B target. Define acceptance criteria, throughput gains vs quality loss, token-level metrics, and failure fallbacks; include cost curves and ablation plan.

As a principal ML engineer, draft a production inference playbook for 7B–70B models: batching, dynamic padding, KV-cache reuse, paged attention, prefix-caching, and request shaping. Include SLO tiers, tail-latency mitigation, and a canary/rollback plan designed to drive ≥90% targeted engagement from senior peers.