Explore Prompts

Design a 2-week study plan covering basic networking, OS fundamentals, and web security (OWASP Top 10). Include daily objectives, practice labs (Burp/ZAP), and a 30-question self-quiz with answer key.

Implement constrained decoding: lexically constrained beam search, regex/JSON constraints, and numerical guards. Provide examples and edge-case tests.

Design a router that selects models/policies per task: classifier gates, uncertainty bands, and backstop abstain. Provide online learning with feedback.

Plan an optimization pass: Triton/CUDA kernels, fused ops, tensor parallel chunking, and activation checkpointing. Provide profiling snapshots and gains.

Produce a latency decomposition (queue→prefill→decode→post). Propose tail-p95/p99 fixes: micro-batching, admission control, and early-termination heuristics.

Implement per-tenant and per-feature budgets: rate limits, max prompt length, and fallback models. Provide business rules and alerts to prevent runaway spend.

Define observability: token usage distributions, tool call success, citation density, and hallucination alerts. Provide redaction-safe logs and dashboards.

Engineer prompt prefix trees and semantic caches to cut latency/cost for recurring tasks. Provide hit-rate models and invalidation policy.

Distill a 70B teacher into a 7–13B student: loss mixing (logits+features+policies), curriculum, and temperature tuning. Provide downstream eval deltas.

Recommend when to use LoRA/QLoRA vs full finetune. Define rank search, target layers, and quantization-aware adapters. Include memory/perf tables per GPU class.

Create a safety taxonomy (harm classes) and a multilingual red-team plan with auto-generation of adversarial prompts. Provide block/transform policies and human review paths.

Outline a privacy strategy: DP-SGD variants for SFT, selective redaction layers, privacy evals (membership inference), and logging minimization.