Fine-Tuning

NCP-GENL · 26 questions

• 311 supervisors labeled preferred versus rejected triage replies and lack full RLHF infrastructure. Which alignment approach fits?
• Permit reviewers have gold “complete application” letters ready. What alignment order should the city follow?
• Housing counselors ranked groups of resource-referral answers from best to worst. Which method can use those group preferences?
• Public-health leaders can staff clinician raters for outbreak SMS tone and accept higher ops cost. When is RLHF the better choice?
• Court self-help drafts sometimes invent filing fees. How should preference data encode that failure?
• Transit customer-service SFT is about to start. What data curation step supports intentional policy compliance?
• After SFT, resident satisfaction with a city assistant has stalled despite more gold demos. What should the team consider next?
• Ordinance retrieval should map paraphrased resident questions to the same code section. Which fine-tuning signal helps?
• A city wants a short-term rental FAQ specialty without duplicating the shared base LLM. Which PEFT choice fits?
• Library branches need children’s desk help and research-desk help without reloading full model weights. What approach fits?
• A floodplain office has only a tiny labeled Q&A set and scarce GPU hours. Which parameter-efficient approach fits best?
• The county assessor wants photo captions and deed snippets for the same parcel to land near each other in embedding space. What training signal should staff use?
• Emergency management needs Spanish and Vietnamese alert packs on one shared decoder backbone. How should the packs ship?
• A water utility freezes its base LLM weights and only trains adapters when the chemical glossary updates. What is the main benefit?
• Permit QA fine-tuning shows falling train loss while validation exact-match plateaus for three consecutive evals. What should the team do?
• A public-health LLM project summarizes outbreak briefs and also classifies triage intents. How should metrics be chosen?
• A clerk chatbot begins quoting memorized party names from training cases. How can early stopping help?
• A housing alignment run must improve preference win-rate without raising hallucination rate. What early-stop rule fits?
• Transit delay-explainer training needs reliable early stopping. How should validation be scheduled?
• A school IEP summarizer improves BLEU but clinician Likert scores fall. What should the team prioritize for stopping and promotion?
• After applying a LoRA adapter, a tax bot invents rebate amounts. What impact assessment action is appropriate?
• Clerk Q&A invents ordinance numbers. Which fine-tuning change best mitigates that hallucination pattern?
• Housing ops will only allow PEFT updates in production. Why does that policy improve safety?
• Public works must prove a pothole FAQ LoRA helps more than it harms. What assessment design fits?
• Emergency alerts invent unverifiable shelter bed counts. Which fine-tuning mitigation targets that failure?
• Parks venue names change overnight and a full SFT cycle will not finish before morning events. What should staff apply?