Concept map: RLHF and DPO fundamentals

Taught and used in

LLM architecture and adaptation

Insights 10

• Keep your hypothesis out of the prompt when asking an LLM to judge evidence
• Rerun safety and refusal tests after every fine-tune, even on benign data
• Avoid letting a model family judge its own outputs
• Locate RLHF failures in feedback, reward model or policy before changing the recipe
• Track a gold metric while optimising against a reward model, and stop before it turns down
• Choose DPO for simplicity, but check for distribution shift in the preference data
• Measure output diversity after preference tuning when variety is part of the product
• Check whether a preference-tuning gain is just longer answers

Projects 6

• Zero-shot time-series foundation models vs strong baselines on post-release monthly data (docs-only)
• LLM-as-judge reliability study: agreement with human pairwise labels, position bias and length bias
• Golden-set regression suite that gates prompt changes for a claims-intake triage assistant
• Summary faithfulness evaluation for legislative bills: human claim labels vs automatic checks
• LoRA/QLoRA adaptation plan for fine-grained banking intents (implementation-ready, docs-only, GPU optional)
• LLM question-answering service: versioned golden set, regression gate and request observability