Explore Prompts

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.

Design a multilingual alignment plan (zh/ja/hi/id/pt/en): shared subword policy, cross-lingual instructions, and locale-specific refusal tuning. Provide leakage checks.

Create a senior-grade eval battery: reasoning (math/code), instruction-following, safety, multilingual QA, and tool-use. Include uncertainty intervals and power analysis for A/Bs.

Define JSON schema contracts with type coercion, partial output recovery, and EBNF constraints. Provide test-time correction and repair strategies.

Specify a lightweight agent loop with decomposition, execution, and critique. Provide termination conditions, trace logging, and loop unroll limits.

Design a tool-use curriculum: function signatures, schema discovery, tool reliability scoring, and retry/backoff policy. Include sandboxing and cost guards.

Create a hallucination detector using entailment+attribution signals. Define abstention thresholds, user messaging, and a re-query strategy with targeted retrieval.