Research on Domain Large Models by Fudan University Knowledge Factory Lab

Background: GPT‑4 marks a turning point, showing strong world knowledge but also high inference cost and limited domain applicability.

Challenges: High inference cost, capability gaps in complex decision‑making, and lack of collaboration with existing enterprise workflows hinder practical adoption.

Domain Adaptation: Discusses data quality and proportion for domain LLMs, introduces source‑enhanced tagging (e.g., “wiki”, “news”, “novel”) to improve reliability, and presents a hierarchical corpus classification scheme for better fine‑tuning.

Capability Enhancement: Focuses on improving instruction following, JSON output, and self‑correction via multi‑step answer generation (PAM), as well as command‑generation correction based on error feedback.

Collaborative Work: Proposes a hybrid workflow where traditional models handle most tasks while LLMs are reserved for open‑world reasoning, knowledge‑base verification, and few‑shot learning; also describes knowledge extraction, alignment, and relation extraction pipelines.

Retrieval‑Augmented Generation (RAG): Combines sparse (BM25) and dense (BGE) retrieval, uses source tags to decide strategy, and enforces provenance by decoding hard constraints with special brackets to ensure quoted text matches source.

Conclusion: Summarizes the research directions for deploying domain‑specific large models in practice.